Timeline: When Do Various Thinkers Expect AGI, ASI, and the Singularity?

Last updated: 2026-05-23

1. Definitions

It helps to begin with terminology, because much of the apparent disagreement in AI timelines is a disagreement about words. The terms below are used inconsistently across sources. For this document, they mean:

• AGI (Artificial General Intelligence) — AI that can perform any intellectual task a human can, at least at human level, across domains. Broadly competent, not narrowly superhuman.
• ASI (Artificial Superintelligence) — AI that substantially exceeds the best human performance in virtually every cognitive domain, including scientific creativity, social intelligence, and general wisdom.
• Singularity — the point at which AI-driven change becomes so rapid and profound that human civilization is fundamentally and irreversibly transformed. Often associated with recursive self-improvement.
• HLMI (High-Level Machine Intelligence) — used in expert surveys to mean machines that accomplish every task better and more cheaply than human workers. Roughly equivalent to AGI.
• FAOL (Full Automation of Labor) — all occupations are fully automatable; machines carry out every task better and more cheaply than humans.

When Altman says “AGI by 2025” and LeCun says “10+ years,” they are partly talking about different things. Much of the spread in timelines reflects this rather than disagreement about the underlying evidence.

This document tracks stated predictions, not the repo’s endorsement of those predictions. New model releases should update the evidence base, but they should not automatically move timeline estimates unless the source’s actual forecast changes.

2. Timeline Overview

3. Kurzweil’s Timeline

Ray Kurzweil’s predictions come primarily from The Singularity is Near (2005) and The Singularity is Nearer (2024). His framework is built on the law of accelerating returns: the observation that computing power per dollar grows exponentially.

The Six Epochs

Kurzweil frames all of history as six epochs of information processing:

1. Physics & Chemistry — the birth of physical laws.
2. Biology — molecules complex enough to encode organisms (DNA).
3. Brains — animals storing and processing information neurally.
4. Technology — humans translating thought into complex action (tools, language, civilization).
5. Merger of Human and Machine Intelligence — brain-computer interfaces directly extending cognition. The current transition, on his account.
6. The Universe Wakes Up — intelligence spreads through the cosmos, turning matter into “computronium.”

Key Milestones (Kurzweil’s Claims)

The Programming Feedback Loop

Kurzweil identifies computer programming as “the main bottleneck for superintelligent AI.” His argument:

“Once we develop AI with enough programming abilities to give itself even more programming skill (whether on its own or with human assistance), there’ll be a positive feedback loop.”

This is the most empirically testable of his claims. The agentic coding trajectory in Section 5 below offers the most direct early evidence — for or against — that this loop is activating.

Assessment

Kurzweil has a mixed track record. He has been strong on exponential trends in computing costs and weaker on specific capability milestones. His timeline is more aggressive than expert survey medians, but increasingly within the range of industry leader estimates. The 2029 Turing test prediction, once considered wildly optimistic, now sits inside the Metaculus community range.

4. Other Predictions

Industry Leaders

Dario Amodei (Anthropic CEO). Amodei has anchored on “powerful AI” arriving in 2026–2027 — intellectual capabilities matching Nobel Prize winners across disciplines. Anthropic’s official position (March 2025) is that “powerful AI systems will emerge in late 2026 or early 2027.” In January 2026, his 20,000-word essay “The Adolescence of Technology” warned that “humanity is about to be handed almost unimaginable power.” In March 2026, speaking at Morgan Stanley, he argued that scaling laws have “not hit a wall at all” and predicted “radical acceleration in 2026.”

Sam Altman (OpenAI CEO). In November 2024, Altman said “we are now confident we know how to build AGI as we have traditionally understood it.” January 2025 brought the claim that GPT-5 was “already smarter than me in many ways,” along with a prediction of superintelligence by 2030. An earlier estimate placed superintelligence by end of 2028: “more of the world’s intellectual capacity could reside inside of data centers than outside.” Corporate actions and releases have matched the rhetoric — the $500B Stargate project, 800M+ weekly ChatGPT users, the $6.5B IO acquisition, GPT-5.3-Codex (February 2026), and GPT-5.5 (April 2026). OpenAI frames GPT-5.5 as a step toward agentic computer work: coding, online research, data analysis, documents, spreadsheets, and operating software across tools. Late-April 2026 partnership changes match the deployment thesis: OpenAI loosened Microsoft exclusivity while remaining tied to Azure, then brought OpenAI models, Codex, and managed-agent workflows into AWS Bedrock environments. The caveat in his own words is worth keeping: Altman acknowledges AGI “has become a very sloppy term.”

Demis Hassabis (Google DeepMind CEO). In March 2025, Hassabis estimated “5 to 10 years” to human-level AI — placing AGI in 2030–2035. At the India AI Impact Summit 2026 he narrowed this to “maybe within the next five years,” conditional on “one or two more major breakthroughs on the level of the Transformer or AlphaGo.” He sets a higher bar than others, requiring genuine invention and creativity (“Could a system invent Go, or come up with relativity?”). He observes that coding and math are progressing fastest, while scientific discovery and creative reasoning remain harder.

Elon Musk (xAI). Musk claims AGI by year-end 2026, and that Grok 5 (6T parameters, Q1 2026) has “~10% chance of achieving AGI.” xAI was acquired by SpaceX at a $250B valuation in February 2026. His track record across domains skews toward aggressive timeline claims.

Anthropic as revealed preference. Anthropic’s stated powerful-AI timeline remains late 2026 / early 2027. Its April 2026 Project Glasswing release posture is strong revealed evidence: Anthropic treats Claude Mythos Preview as useful enough for critical defensive cybersecurity but too risky for general availability. Claude Opus 4.7, released April 2026, is positioned as the broadly available model for advanced software engineering and multi-step work while cyber safeguards are tested before Mythos-class models are released more widely.

OpenAI as revealed preference. GPT-5.3-Codex and GPT-5.5 both emphasize long-running, tool-using professional work rather than chat alone. OpenAI reports internal use of Codex/GPT-5.5 to debug training, deployment, evaluation, and serving infrastructure. Anthropic’s Claude Code product page makes a similar claim (“the majority of code at Anthropic is now written by Claude Code”) and lists named enterprise migrations (Stripe Scala-to-Java in 4 days, Wiz 50,000-line Python-to-Go in ~20 hours of active dev time, Rakuten new-feature delivery 24 to 5 working days). Redwood Research’s “Is 90% of code at Anthropic being written by AIs?” rebuttal is the right calibration here: the most defensible sub-metric is “lines of code merged” — likely a majority — while self-reported productivity gains remain around 20–40%. The vendor claims are best read as revealed-preference compounding evidence, not as audited engineering-productivity multipliers. This is weak evidence for Kurzweil’s programming feedback loop, but not yet evidence of autonomous recursive self-improvement.

Ilya Sutskever (Safe Superintelligence Inc.). The contrarian position. Sutskever argues “the age of simple scaling is ending” and that the next breakthrough will require fundamentally new learning methods. He is running SSI at a $32B valuation with roughly 20 employees and zero revenue — itself a kind of revealed preference.

Jensen Huang (Nvidia CEO). In March 2024, Huang predicted AGI within 5 years (2029), defined as passing a set of tests 8% better than most people. By November 2025, he had shifted to claiming AGI is “already here today” for practical applications. The definitional shift is itself informative.

Yann LeCun (Meta Chief AI Scientist). “At least a decade and probably much more.” He cautions readers that “if someone claims AGI is just around the corner, do not believe them.” His core argument: LLMs will not lead to AGI; fundamentally new architectures (world models, JEPA) are required. He estimates 3–5 years for new architectures to mature, but treats those as prerequisites for AGI rather than AGI itself.

Academic Views

Nick Bostrom does not give specific AGI dates. He has argued that a 10% probability of HLMI failing to arrive by 2075 or even 2100 “seems too low.” His focus is on existential risk given any timeline, not on prediction.

Max Tegmark (MIT) observes that “AI capabilities have advanced so rapidly that they’ve collapsed previous predictions about AGI timelines.” He does not commit to specific dates but emphasizes urgency.

Stuart Russell (UC Berkeley) notes that “everyone has gone from 30–50 years to 3–5 year timelines.” He assigns roughly a 50% chance that AGI stalls, and a 30% chance it can be built under the present paradigm. His focus is safety rather than timing.

Survey Data

ESPAI 2023 (AI Impacts, 1,714 AI researchers) found a 50% chance of HLMI by 2047 (down from 2060 in 2022), and a 50% chance of FAOL by 2116 (down from 2164 in 2022). The 13-year shift in a single year reflects the impact of GPT-4 and similar advances on expert opinion.

2024–2025 forecaster surveys are more aggressive: 50% expect HLMI by 2030, 90% by 2040. At the same time, 76% of AI experts think it unlikely that current LLM methods alone can scale to AGI.

Metaculus (Feb 2026) estimates a 25% chance of AGI by 2029 and a 50% chance by November 2033. In four years, the mean estimate dropped from 50 years to 5 years. The notable detail: forecasts moved outward during late 2025, from July 2031 to November 2033.

Forecast aggregates (May 2026). Goodheart Labs’ dashboard, aggregating Metaculus, Manifold, and Kalshi, estimated AGI in 2031 as of May 23, 2026. Its 80% interval was 2027–2043, which captures the state of the field better than any single point estimate: near-term AGI remains plausible in markets and forecasts, but the uncertainty band still spans more than a decade. The dashboard is a synthesis rather than a primary forecast with one clean definition. It combines differently worded questions, including Metaculus-style AGI questions and Turing-test markets.

Frontier Release and Deployment Evidence (April–May 2026)

The April–May 2026 frontier releases and deployment changes do not settle the AGI timeline. They do update the evidence in six directions worth tracking.

1. Shorter timelines get stronger evidence from autonomy. GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro / Deep Research Max, and Claude Mythos Preview all point toward models that can plan, use tools, manage context, and verify work over longer horizons.
2. Longer timelines get stronger evidence from deployment friction. Labs are gating models, adding safeguards, limiting cyber/bio access, and emphasizing trusted access rather than broad release. Capability is arriving faster than institutional deployment readiness.
3. Definitions matter more than ever. If AGI means “economically useful autonomous computer work,” the frontier is moving quickly. If it means robust general invention, stable agency, and reliable real-world autonomy across domains, the evidence is still incomplete.
4. Deployment architecture is now timeline evidence. OpenAI’s multicloud movement, AWS managed-agent integration, and U.S. classified-network AI agreements show that frontier systems are moving into enterprise and national-security infrastructure before broad AGI is resolved as a concept.
5. Google’s Gemini 3.5 release strengthens the “agents as path to AGI” frame. Gemini 3.5 Flash is explicitly framed around long-horizon agentic workflows, coding, subagents, and broad distribution through Search, Gemini, Antigravity, Android Studio, and enterprise platforms. This does not prove AGI is near, but it makes agentic workflow length a better leading indicator than chat intelligence alone.
6. On-prem and governed-data deployment matter. The OpenAI-Dell Codex partnership, Microsoft/EY implementation work, and NVIDIA’s new hyperscale / ACIE / edge reporting categories all point in the same direction: timelines depend on whether agents can operate inside real organizational data, infrastructure, and permission boundaries.

New Research Signals (May 2026)

Three May 2026 research signals sharpen the timeline picture without moving the central forecast by themselves.

1. Agentic systems as an AGI pathway. Liao et al. argue that AGI may emerge from agentic systems rather than monolithic model scaling, because routing, multi-agent composition, and DAG-style task structures can improve generalization and sample efficiency. This supports the repo’s focus on agents and software-development milestones as leading indicators.
2. Domain transfer is still brittle. Phoenix-bench shows software-engineering agents lose 37–58% when moved from SWE-bench Verified-style software tasks to realistic hardware-debugging tasks. That is a concrete warning against reading high coding scores as broad engineering generality.
3. AGI remains definitionally unstable. Fletcher and Vu Khan emphasize that AGI is conceptually and socio-technically problematic, and that pathways differ across frontier proprietary models, open-weight models, domain-specific systems, and sovereign model strategies. This reinforces this document’s core caution: timelines are partly forecasts about definitions, not only capabilities.

Other Individual Estimates

• Shane Legg (DeepMind co-founder): 50% chance of “minimal AGI” by 2028.
• Andrew Critch (AI researcher): 45% chance of AGI by end of 2026.
• Matthew Barnett (training loss extrapolation): median for transformative AI ~2033.
• Daniel Kokotajlo / AI 2027 project: month-by-month AGI by 2027, ASI shortly after. Self-assessment (early 2026): progress at ~65% of predicted pace; median shifted from 2028 to 2029.

5. Software Development Milestones

The most concrete near-term trajectory runs through software development, where AI capabilities are most measurable. Based on current progress extrapolation:

These milestones track the Kurzweil programming feedback loop in real time. The 2025 milestone (agentic coding) is realized. The 2026 infrastructure milestone is now shipping across frontier providers: OpenAI/AWS Bedrock Managed Agents and the May 18, 2026 Bedrock Stateful Runtime; Claude Managed Agents with self-hosted sandboxes and MCP tunnels; Gemini / Antigravity 2.0 agent surfaces and Managed Agents in the Gemini API; the Cursor SDK with Composer 2.5 and Jira integration; Devin 2.x release features; GitHub Copilot Cloud Agent across IDEs running Opus 4.7 and GPT-5.5; and Codex moving toward governed hybrid and on-prem environments. Capital markets price this trajectory aggressively: Cognition (the maker of Devin) is reportedly targeting a $25B raise, roughly 2.5× its September 2025 valuation. The capability milestone is weaker than the infrastructure one. Vendor case studies show large refactors (Stripe, Wiz, Rakuten, Goldman Sachs, Visma), but independent evaluations (Answer.AI on Devin 1.0, METR scaffolding work, Phoenix-bench, Khanal et al. on long-horizon reliability) continue to show brittle reliability and heavy human orchestration.

The most important update from the new software-development report is therefore not “coding agents are solved.” It is narrower and more useful: the agent runtime substrate is arriving faster than audited autonomous capability. The next leading indicator is whether agents can keep identity, memory, tests, permissions, and regression accountability across multi-week projects — not whether they can produce another impressive isolated patch.

Recursive-coding evidence in production is also accumulating, though it is not yet runaway self-improvement. DeepMind’s May 7, 2026 AlphaEvolve impact post reports concrete deployed wins (~30% variant-detection error reduction for PacBio sequencers via DeepConsensus; AC Optimal Power Flow GNN feasibility from 14% to >88%; ~10x lower error on Willow quantum circuits), extending the May 2025 results on a 23% Gemini training matmul speedup and a 48-multiplication 4x4 complex matmul beating Strassen. Sakana’s Darwin Godel Machine reports SWE-bench 20.0% to 50.0% and Polyglot 14.2% to 30.7% through open-ended agent self-modification. Jeff Clune’s new startup Recursive has reportedly raised $650M at a $4.65B valuation aimed explicitly at this pipeline. IEEE Spectrum’s May 2026 framing remains the honest synthesis: recursive self-improvement is “emerging, but humans are still in the loop.”

By 2030, if this trajectory holds, the boundary between “tool” and “autonomous agent” in software development will be substantially blurred — though likely still within human-defined goal structures.

6. Constraints and Skepticism

Several reasons timelines might slip are worth holding in mind.

Definitional inflation. As AI achieves previous AGI benchmarks, the goalposts move. What counts as “AGI” keeps shifting upward.

Scaling limits. Training costs rise 10–50× per generation. Energy, data quality, and memory are real constraints.

LLM plateau. 76% of AI experts doubt LLMs alone can reach AGI. If new architectures are needed, add years for R&D and deployment.

Integration versus capability. Demonstrating AGI-level performance on benchmarks is different from deploying reliable, safe systems in the real world.

Benchmark saturation and contamination. OpenAI now recommends against using SWE-bench Verified for frontier launches because of flawed tests and training-data exposure. SWE-bench Pro, live benchmarks, and task-time evaluations carry more signal than 85%+ Verified scores.

Long-horizon reliability decay. METR’s time-horizon work, tau-bench, and new reliability-science papers all point to the same issue: agents can look strong on one-shot task success while repeated, longer, or more interactive workflows decay sharply.

Agent domain transfer. Agents that work in software repositories may fail in hardware, physical systems, biology, law, or operations, because each domain has different feedback loops, observability, and verification structure.

Skill and tool supply-chain risk. Agent capability increasingly depends on external skills, tools, plugins, MCP servers, and registries. If those artifacts are misdescribed or malicious, longer-horizon autonomy becomes less reliable and harder to govern.

2025 forecast recession. Metaculus forecasts moved outward in late 2025, suggesting earlier optimism may have been premature.

Critical bottleneck window. 80,000 Hours identifies 2028–2032 as a likely bottleneck period — either AGI arrives within five years of 2026, or progress slows significantly.

Historical pattern. AI has a long history of over-promising on timelines: Minsky’s “within a generation” in 1967, expert systems in the 1980s, and so on. This sounded convincing at the time. History, in each case, did not cooperate.

The divergence between industry leaders (2026–2029) and expert surveys (2047) is itself a data point. Industry leaders have commercial incentives to project optimism. Survey respondents have recency bias. Neither is obviously more reliable.

7. Our Assessment

The baseline model in BASELINE.md sits in the moderate-acceleration camp. The reasoning is as follows.

Software development milestones (2025–2030) are the most credible near-term predictions because they are already partially observable. They are being tracked actively here.

Narrow AGI-like capabilities in specific domains — coding, math, certain sciences — by 2028–2030 appear plausible.

Broad AGI — the kind that matters for Singularity discussions — is harder to pin down. The 2031–2033 aggregate / Metaculus range remains more defensible than either the industry-optimistic (2026–2027) or the survey-conservative (2047) estimates, although April–May 2026 releases and deployment changes slightly strengthen the case for earlier capability diffusion.

ASI and Singularity remain too speculative for meaningful point estimates. Kurzweil’s 2045 Singularity is a useful boundary marker, not a prediction we’d stake the model on.

The programming feedback loop is the single most important thing to watch. If agentic coding capabilities compound faster than the year-by-year trajectory in Section 5, timelines compress. If they plateau, they extend.

Agentic distribution is now the key leading indicator. The question is not just whether the next model is smarter. It is whether agents can sustain useful work across tools, teams, permissions, and domains without accumulated errors dominating. The May 2026 evidence strengthens the infrastructure half of this claim. METR, SWE-bench Pro, Phoenix-bench, tau-bench, and Answer.AI-style evaluations keep the capability claim bounded.

Benchmark-saturation caveat. SWE-bench Verified is now consensus-contaminated (OpenAI, February 23, 2026). Frontier scores there cluster at 85–94%, while the same models score in the 23–65% range on SWE-bench Pro — a ~40–50pp gap for the same model. Terminal-Bench 2.x top-cluster differences fall inside ±1.5–2.6pp confidence intervals and depend heavily on the agent harness, not the model. METR’s task suite has saturated past 16 hours: the May 8, 2026 page note says “measurements above 16 hrs are unreliable with our current task suite,” and Claude Mythos Preview’s 95% confidence interval on the 50%-time horizon is 8.5–55 hours. Treat any single benchmark number above 80% with suspicion until SWE-bench Pro, the next-gen METR suite, and harness-normalised leaderboards stabilise.

The honest summary: nobody knows. The spread from “already here” (Huang) to “at least a decade” (LeCun) to “50% by 2047” (ESPAI) reflects genuine uncertainty, not just disagreement. The approach taken here is to track leading indicators — especially software development milestones — update quarterly, and resist both hype and complacency.

Sources: Kurzweil (2024), ESPAI 2023 survey, Metaculus community forecasts (Feb 2026), Goodheart Labs forecast aggregate (May 2026), 80,000 Hours (2025), Kokotajlo/AI 2027 (2025), public statements from Amodei, Altman, Hassabis, Huang, Musk, Sutskever, LeCun, Russell, Tegmark, Bostrom, April-May 2026 model releases from OpenAI, Anthropic, and Google, May 2026 agent research, and April-May 2026 deployment evidence from OpenAI, Microsoft, AWS, Dell, NVIDIA, Google, and U.S. classified-network AI agreements. See _data/sources.yml for the full registry.