When Will AGI Arrive?

Key Links

Perhaps the most consequential forecasting question in history: When will we develop AI systems that match or exceed human-level intelligence across virtually all domains?

The answer determines how much time we have to solve alignment, whether to prioritize AI safety over other causes, and how urgently we need governance frameworks.

Defining AGI

The challenge: No consensus definition of AGI

Common criteria:

• Can perform any intellectual task humans can
• Can learn new tasks quickly with minimal data
• Generalizes broadly across domains
• Autonomous planning and goal-pursuit
• Economic productivity matching human workers

Proxy metrics:

• Pass rigorous expert-level tests across domains
• Outperform median human on most economically valuable tasks
• Can do the job of an AI researcher (recursive self-improvement)
• $100B+ annual economic value

Note: Debate conflates different concepts:

• Human-level AI (matches median human)
• Transformative AI (drastically changes world)
• Artificial General Intelligence (truly general intelligence)
• Superintelligence (exceeds all humans)

Timeline Camps

Key Forecasts and Positions

Key Cruxes

What Would Update Timelines?

Evidence for shorter timelines:

• GPT-5/6 showing qualitative leap in reasoning and planningCapabilityReasoning and PlanningComprehensive survey tracking reasoning model progress from 2022 CoT to late 2025, documenting dramatic capability gains (GPT-5.2: 100% AIME, 52.9% ARC-AGI-2, 40.3% FrontierMath) alongside critical...Quality: 65/100
• Successful scaling past data limits
• AI substantially accelerating AI research
• Solving ARC benchmark or similar generalization tests
• Continued exponential capability gains

Evidence for longer timelines:

• Scaling 100x with only incremental improvements
• Hitting hard data or compute walls
• Persistent failures on key capabilities despite scale
• Need for architectural breakthroughs that don't arrive
• Progress slowing on key benchmarks

Historical Track Record

Past AGI predictions:

• 1958: "Machines will be capable, within twenty years, of doing any work that a man can do" - Herbert Simon
• 1965: "Machines will be capable, within twenty years, of doing any work that a man can do" - Herbert Simon (updated)
• 1970: "In from three to eight years we will have a machine with the general intelligence of an average human being" - Marvin Minsky
• 1980s: Expert systems will lead to AGI by 2000
• 2000s: AGI by 2020

Pattern: Always 20-30 years away. Should we believe this time is different?

Arguments it's different now:

• Have empirical scaling laws, not just speculation
• Concrete progress on benchmarks and capabilities
• Massive investment and resources
• Clear path forward (scaling) vs unknown unknowns

Arguments it's the same:

• Still don't understand intelligence
• Benchmarks may not capture true intelligence
• Economic and technical obstacles remain
• Same overconfidence as past predictions

The Distribution Shape

Most forecasters have heavy-tailed distributions:

Short tail (optimistic):

• 5-10% chance: AGI by 2027
• 20-25% chance: AGI by 2030
• Driven by: Scaling working, rapid progress, no blockers

Central mass:

• 50% chance: AGI by 2035-2040
• Most likely scenario: Continued progress with some obstacles

Long tail (pessimistic):

• 20-30% chance: AGI after 2050
• 5-10% chance: Never with current paradigms
• Driven by: Fundamental limits, need for new paradigms

Wide uncertainty is rational given:

• Deep uncertainty about scaling limits
• Unknown unknowns
• Dependence on definition
• Historical poor track record

Implications for Different Timelines

If AGI by 2027-2030:

• Extremely urgent to solve alignment NOW
• Current safety research may be too slow
• Need immediate governance action
• Race dynamics critical concern
• May not get warning signs

If AGI by 2030-2040:

• Time to iterate on safety
• Can learn from weaker systems
• Governance frameworks can develop
• Safety research can mature
• More coordination opportunities

If AGI after 2050:

• Safety research can be thorough
• Governance can be careful and democratic
• Current hype may be overblown
• Other causes may be higher priority
• Different paradigms may emerge

Economic vs Philosophical AGI

Important distinction often blurred:

Economically transformative AI:

• Automates most jobs
• Generates trillions in value
• Fundamentally changes society
• Might come soon (2027-2035)
• Doesn't require "general" intelligence

Philosophically general intelligence:

• True understanding across all domains
• Quick learning like humans
• Causal reasoning and abstraction
• Might require paradigm shifts
• Could be much further (2040+)

Why it matters:

• Economic transformation could happen without "AGI"
• Most impacts come from economic transformation
• But existential risk might require true AGI
• Definitions determine timeline estimates

The Compute Bottleneck

Different views on compute as limiting factor:

Optimistic: Compute is abundant

• Moore's law continues
• Efficiency improvements ongoing
• Cloud compute scales easily
• No physical limits near

Pessimistic: Compute limits soon

• Training costs becoming prohibitive ($1B+)
• Energy and chip constraints
• Economic feasibility limits
• Can't scale 1000x more

Resolution matters:

• If compute limits: Longer timelines, regulated by economics
• If compute abundant: Timelines depend on algorithmic progress

The China Factor

How does China affect timelines?

Arguments China accelerates:

• Competition drives urgency
• Massive investment
• Less safety caution
• Different approaches might work

Arguments China doesn't change much:

• US still ahead on capabilities
• Chinese models lag 1-2 years
• Limited to similar approaches
• Compute restrictions bite

Strategic implications:

• If China racing: Pressure for short timelinesAnalysisShort AI Timeline Policy ImplicationsAnalyzes how AI policy priorities shift under 1-5 year timelines to transformative AI, arguing that interventions requiring <2 years (lab safety practices, compute monitoring, emergency coordinatio...Quality: 62/100
• If US leads comfortably: Can afford to be cautious
• Matters for regulation and safety investment

Recursive Self-Improvement

Wild card: AI accelerating its own development

If happens soon:

• Could dramatically shorten timelines
• "Singularity" scenario
• Hard to predict outcomes
• Very fast takeoff possible

If doesn't happen:

• Progress continues at current pace
• More time to prepare
• Gradual development allows adjustment

Current status:

• AI assists with coding and research
• But not yet transformative acceleration
• Unclear if/when recursive improvement kicks in

Base Rates and Reference Classes

What should we compare to?

Reference class: Major technologies

• Electricity: 50 years from invention to transformation
• Computers: 40 years from invention to ubiquity
• Internet: 20 years from invention to transformation
• Suggests: Long timelines (decades)

Reference class: Exponential technologies

• Semiconductors: Exponential for 50+ years
• Genomics: Exponential progress continues
• Suggests: Continued rapid progress possible

Reference class: Breakthroughs

• Manhattan Project: 3 years when focused
• Apollo Program: 8 years with resources
• Suggests: Massive resources can compress timelines

Problem: AGI is unique, unclear which reference class applies