How Close Are We to AGI? Live Timeline & Expert Predictions

What is AGI

The difference between today's AI and AGI

Most people have used ChatGPT or Claude. That's not AGI. Here's the actual gap.

🤖

Today's AI (narrow)

Extremely capable at specific tasks it was trained on — writing, code, maths, image recognition. Give it a task outside its training and performance drops sharply. It cannot learn from experience, set its own goals, or operate reliably without human oversight.

🧠

AGI (what we're tracking)

An AI that can do any intellectual task a human can — across all domains, adapting to things it's never seen before. It can set its own goals, learn continuously from new experiences, and operate without needing humans to define every step.

The gap between these two things is what this page tracks

Why there's no single answer: "AGI" means different things. Some define it as replacing the median knowledge worker (might arrive sooner). Others define it as matching all human cognitive abilities including physical tasks and genuine creativity (much harder). The 2031 estimate is for the first definition — economic AGI. Full general intelligence is likely later.

Capability signals

What AI can already do — and what's still missing

Each bar shows how close current frontier AI is to matching a human expert at that skill. Green = solved. Red = the remaining gap.

Essentially solved — AI at or near human expert level

Language & reasoning

91%

Mathematics (competition)

88%

Scientific knowledge

74%

Code generation

72%

Progressing — significant gaps remain

Multimodal understanding

55%

Visual spatial reasoning

51%

Multi-week autonomous tasks

44%

Hard unsolved — these are the real bottlenecks

Persistent memory & learning

35%

Open-ended novel research

38%

Physical world understanding

28%

Self-improvement

22%

The global race

Which countries are leading the AGI race?

AGI won't be built by one company. It'll be built by whoever can stack the most frontier labs, compute, and talent — and right now that race is concentrated in two countries.

🇺🇸 United States

Home of every frontier model lab

Overall lead score

Strong

Frontier labs

95%

AI compute share

~70%

AI research papers

~40%

Gov't AI investment

$500B+

OpenAI, Anthropic, Google DeepMind, Meta AI, xAI — all US-based. The US holds a commanding lead in frontier model capability and private investment. The Stargate initiative committed $500B in AI infrastructure through 2030.

🇨🇳 China

Closing the gap faster than most expected

Overall lead score

Closing

Frontier labs

35%

AI compute share

~15%

AI research papers

~38%

Gov't AI investment

$150B+

DeepSeek shocked the world in early 2025 with a model rivalling GPT-4 built at a fraction of the cost. Baidu, Alibaba, and Huawei are all active. China's chip constraints (US export controls on Nvidia) are the primary limiting factor — but China is aggressively building domestic alternatives.

🌍 Europe + Rest of World

Regulation-first, capability-second

Overall lead score

Behind

Frontier labs

<10%

AI research papers

~22%

Regulatory activity

High

Mistral (France) is the only European lab producing competitive frontier models. The EU AI Act is the world's most comprehensive AI regulation framework — but has led some researchers to argue Europe is trading capability for compliance. UK, Canada, and Israel have notable research presence but no frontier labs.

The chip chokepoint: The most important single variable in this race is Nvidia GPU supply. The US controls Nvidia's export licenses. China's compute access is constrained. If China closes the chip gap through domestic production (SMIC, Huawei), the race gets significantly tighter.

Milestone timeline

The moments that moved the forecast

Every major AI breakthrough in recent years has pushed the expert prediction earlier. Here's why.

2017

Transformer architecture invented at Google

The paper "Attention Is All You Need" changed everything. Every powerful AI system today — GPT-4, Claude, Gemini — is built on this. Before this, AI had a memory problem. After it, scale became the strategy.

2020

GPT-3 shows AI can generalise

For the first time, a single model could write poetry, answer legal questions, and debug code — without being specifically trained for any of them. Expert predictions shifted from "maybe never" to "probably this century."

2022

ChatGPT: 100 million users in 60 days

The fastest product adoption in history. This wasn't just a technology milestone — it was a signal that AI was ready for general use. Forecasts shifted again: the median prediction moved from 2055 to 2048.

2023

GPT-4 passes the bar exam — top 10%

Not a party trick. A model that passes a professional exam at expert level is performing at the level of a trained human in a high-stakes domain. The "Sparks of AGI" paper from Microsoft triggered serious academic debate.

2024

o1 reasoning: AI learns to think before it speaks

OpenAI's o1 model showed that giving AI time to reason through a problem — before answering — dramatically improved accuracy on hard problems. Performance on PhD-level science benchmarks exceeded most human experts.

Now — Aug 2026

AI completing multi-day software projects autonomously

Claude Code, Devin, and similar agentic systems are completing entire software projects without step-by-step human instruction. AI is beginning to assist in novel scientific discovery. The forecast median sits at 2031 — down from 2060 in 2019.

~2027–28 (predicted)

AI researchers working at human expert level

Anthropic's internal framing and OpenAI's published roadmaps both point to this window for AI that can autonomously conduct scientific research. Not full AGI — but the line starts to blur.

~2031 (median forecast)

Economic AGI threshold

The Metaculus community and multiple academic surveys converge around this date for AI that can substitute for the majority of remote knowledge work. Uncertainty range is wide: optimists say earlier, sceptics say later.

Expert predictions

What the people building it say

The range is wide — but it's collapsing. In 2020, most serious researchers said "decades away." In 2026, most say "this decade."

2026–27

Dario Amodei

CEO, Anthropic

Calls it "powerful AI" — systems capable of autonomous research at human level. Distinguishes this from full AGI but says it's arriving very soon.

2029

Ray Kurzweil

Futurist, Google

Predicted this date since 2005 and has held it. Defines AGI as passing a two-hour Turing test convincingly across all domains.

2031

Metaculus median

5,000+ forecasters

Community prediction market aggregate. Has moved forward by 4+ years since 2022. Represents the broadest consensus available.

2030s

Demis Hassabis

CEO, Google DeepMind

Most measured of the major lab leaders. Points to remaining gaps in physical grounding and genuine causal reasoning.

Uncertain

Sam Altman

CEO, OpenAI

Has said AGI is "very soon" but avoids a year. OpenAI defines AGI as "AI that outperforms humans at most economically valuable tasks."

2041+

Yoshua Bengio

Turing Award recipient

The most cautious voice among top researchers. Argues current architectures cannot produce genuine understanding or causal reasoning.

Common questions

Frequently asked

AGI stands for Artificial General Intelligence — an AI that can perform any intellectual task a human can, across all domains, without being specifically trained for each one. It can set goals, learn from experience, reason about things it's never seen before, and operate without step-by-step human instruction. Today's AI systems — including ChatGPT, Claude, and Gemini — are extremely powerful but are not AGI. They excel at the tasks they were trained on but struggle with genuine novelty, long-horizon planning, and continuous learning. AGI is the step beyond that.

Because AI has improved faster than almost anyone expected. In 2019, the median expert forecast for AGI was around 2060. After GPT-3 in 2020, it moved to 2055. After ChatGPT in 2022, it moved again. After GPT-4 in 2023, it moved further. Each time a major capability threshold gets crossed earlier than predicted, the entire forecast distribution shifts. The compression from 2060 to 2031 — a 29-year shift in just 7 years — is itself the strongest signal that something unusual is happening with the pace of progress.

Yes to both. The primary acceleration risk is AI-assisted AI research — if frontier models become capable of meaningfully improving their own training or architecture, timelines could compress non-linearly and very fast. The primary delay risks are the data wall (high-quality training data is finite, and synthetic data hasn't fully solved this), and the possibility that planning and physical grounding require architectural breakthroughs rather than just more compute. The uncertainty range is genuinely wide — 2028 to 2040 is a defensible range depending on which definition and which risks materialise.

No. AGI means matching human-level intelligence across all domains. Superintelligence means exceeding human intelligence — potentially by a large margin and across all cognitive tasks simultaneously. Most researchers expect AGI to come first, followed by a period of rapid capability improvement that might or might not lead to superintelligence. The timeline for superintelligence — if AGI arrives — is one of the most contested and consequential open questions in AI research.

The main chart shows the Metaculus community median forecast at different points in time, supplemented by aggregated data from academic AI survey papers. The forecast history is sourced from published Metaculus data, AI Impacts surveys, and structured expert polls. The page is reviewed weekly — when something material changes (a major benchmark result, a significant model release, or a meaningful shift in the Metaculus median), the date and the chart are updated. Each change is logged in the page's last-modified date.

The four hardest remaining problems: (1) Persistent memory — AI cannot learn from new experiences without being retrained from scratch; (2) Long-horizon planning — current models struggle to pursue goals across extended time horizons with many interdependent steps; (3) Physical grounding — AI has no embodied experience of the physical world, which limits reasoning about everyday situations; (4) Self-improvement — the ability of an AI to meaningfully enhance its own capabilities remains extremely limited. These are not just "more of the same" challenges — they may require architectural breakthroughs rather than just scaling up existing approaches.