Artificial general intelligence (AGI) means AI able to generalize across many cognitive tasks, including unfamiliar problems—not just excel at one benchmark. No arrival date is independently verified. Cortist’s provisional, unscored F-001 forecast assigns a 38% chance of independently verified AGI by 31 December 2030.
An editorial judgment, not a scientific measurement or a currently scored forecast.
Cumulative chance by year-end · Subjective model v0.1 · Method ↗
Start with the definition and evidence threshold. Explore the provisional forecast, original record and dated research below.
Artificial general intelligence means broad, transferable AI capability across different cognitive tasks, including unfamiliar ones. There is no universally accepted threshold; one high benchmark score does not establish AGI.
Explore the full AGI guide →No independently verified arrival date exists. Cortist’s F-001 v0.1 estimates a provisional 38% chance by 31 December 2030. This is an editorial judgment, not an observed frequency or scored probability.
Inspect forecast F-001 →AGI refers to broad, general competence. Artificial superintelligence (ASI) would substantially surpass leading human experts across many intellectual domains. They require different evidential thresholds.
Compare AGI and ASI →Independent, repeatable tests would need to assess breadth, unfamiliar-task transfer, and sustained performance under clear criteria. No universal AGI test exists, and Cortist’s F-001 criteria remain under review.
Read the verification protocol →Guide clarified 3 November 2026 · Forecast v0.1 recorded 22 October 2026, unchanged · How Cortist evaluates evidence
Future probability changes require a new dated version and an evidence-based explanation; old records remain available.
One recorded draft (v0.1). No newer probability has been published. The original stays available.
Expert opinions alone do not automatically change a Cortist forecast.Artificial general intelligence (AGI) describes a proposed AI capability that generalizes across many cognitive tasks and unfamiliar problems. There is no universally accepted verification test or independently confirmed arrival date. Cortist’s F-001 assigns a provisional subjective 38% cumulative chance by the end of 2030; this is not an observed or independently scored probability.
Evidence Pulse 2.0 checks the published Cortist News index when JavaScript is available. The source-linked, dated research articles remain accessible through Cortist News.
Latest topic-linked reporting, compared with the separately versioned Cortist assessment. News does not automatically update a forecast.
Evidence interpretation: Developer-agent progress and published expert timelines keep the general-intelligence question active.
What remains unproven: Expert expectations and isolated benchmark successes are not independent proof of AGI.
What to verify next: Independent broad-transfer and unfamiliar-task evaluations, not developer-selected demos.
Snapshot of 2 linked reports. Context is not independent proof or a measured probability change. Dated evidence index · Original forecast v0.1.
Artificial general intelligence (AGI) generally refers to AI that can perform well across a broad range of cognitive tasks, including unfamiliar ones, rather than excelling in a single benchmark. There is no universally accepted pass/fail test. Google DeepMind’s Levels of AGI framework separates breadth of capability from depth of performance and also considers autonomy. Under that approach, solving a difficult mathematics problem or producing convincing prose is evidence about a particular ability; neither alone establishes general intelligence.
A credible test would specify the task distribution, baseline human performance, allowed tools, training-data leakage checks, novelty of problems and the level of autonomy before evaluation. It should cover transfer learning, sustained reasoning, memory, planning and error correction. Public demonstrations deserve attention, but benchmarks selected by a developer cannot be treated as universal proof.
Advances in reasoning, software engineering and multimodal perception show that frontier models are improving on many dimensions. But narrow success on tests does not automatically translate into reliable, independent performance across the range of situations people encounter. Google DeepMind’s March 2026 cognitive taxonomy identifies ten relevant abilities and explicitly calls for better evaluations. It is a measurement proposal, not an announcement that AGI has arrived.
METR’s time-horizon research offers a different lens: the length of software tasks an agent can complete with a specified success probability. It is valuable because extended tasks reveal compounded failures. However, the task suite is predominantly technical, the horizon is based on human expert completion time, and neither a rising horizon nor an extrapolated trend is a direct AGI arrival test.
The responsible answer is a range of competing possibilities, not a countdown. Researchers, companies and forecasting communities use different AGI definitions and provide different horizons. A statement about “human-level AI,” a predicted date for economically useful agents and a claim about general human cognitive competence cannot be placed on one timeline without the original criteria. Cortist preserves dated external claims separately from its own estimate.
F-001 is a cumulative subjective view: 38% by the end of 2030 means the assigned chance of the relevant event having occurred by that date, not a 38% chance in 2030 alone. The original v0.1 curve and year selector remain available on this page. Criteria are still under review; the estimate is not locked, independently verified or scored.
The strongest positive signal would be a pre-registered and independently administered evaluation showing robust performance across unfamiliar domains, sustained autonomy, reliable tool use and clear human comparison groups. Repeated results across organizations and failure audits matter more than a single headline score. Public documentation should specify model version, allowable assistance and how tasks were selected.
Reasons for caution include benchmark contamination, excessive scaffolding, unnoticed human intervention, fragile multi-step execution and definitions that move after a result is observed. Clear negative replications or robust evidence of a capability plateau would also matter. The forecast should change only through a dated, explained new version, never by quietly rewriting its historical curve.
Fluent conversation can demonstrate language competence, but it does not establish how reliably a system can solve unfamiliar problems or learn new rules in an unknown environment. An AI might perform extremely well on a coding test and still struggle with strategic planning, complex real-world constraints or transferring knowledge between subjects. Generality requires evidence across a distribution of tasks, not one impressive output.
Performance must be interpreted together with the amount of help provided. Tool access, human corrections, search and repeated attempts can make a system more useful while complicating comparisons. A trustworthy AGI claim identifies the exact model, allowed tools, task selection, evaluation rules and human baseline.
Narrow AI excels in specific tasks. Today’s general-purpose frontier systems may handle many different tasks but show uneven reliability, especially when problems are novel or multi-step. AGI is a proposed broader adaptability threshold under explicit verification criteria. Artificial superintelligence describes the additional, hypothetical possibility of systems exceeding the best human capability across many intellectual fields.
These are different categories of evidence. Excellence at a single game is not AGI. A frontier model doing useful research is not automatically ASI. The Levels of AGI research framework helps distinguish depth of performance from breadth, autonomy and deployment considerations.
| Category | What it shows | What remains unproven |
|---|---|---|
| Specialist AI | Performance in a defined task | Generality outside that task |
| Frontier general-purpose AI | Broad, sometimes uneven tested ability | Reliable transfer across all new settings |
| AGI (proposed) | Adaptable competence across diverse tasks | An agreed independently resolved event |
| ASI (hypothetical) | Broad performance far beyond leading humans | Inevitable arrival or a certain date |
A serious evaluation would include tasks developed or held out independently, multiple domains, new problems and repeated trials. The test should define success before running the model, document any human intervention and check whether information from the test might already be represented in training material. Evaluators should compare results with suitable humans and disclose uncertainty, cost and failure modes.
This is why a few spectacular benchmark records are insufficient. A benchmark may be valid within its intended scope but not answer whether a machine can work independently across many unfamiliar settings. Outside replication and fresh test construction strengthen claims. The 2026 Google DeepMind cognitive taxonomy is an effort to improve measurement, not a declaration of AGI.
Long projects require an AI system to remember constraints, plan steps, choose tools, recover from errors and verify final outcomes. Even if it performs well on isolated questions, small errors may compound throughout a complex workflow. Reliable completion across longer tasks is therefore valuable evidence about practical capability.
METR’s task-completion time horizon is measured using the duration a human expert needs for tasks of a given difficulty, at a specified AI success probability. It does not indicate that an agent ran without assistance for that many hours, nor does it test every occupation or cognitive domain. It is one signal within a wider AGI assessment.
Progress in reasoning, memory, tool use, planning, multimodal learning, research automation and verification could move systems toward broader competence. Yet there is no established formula that translates more computing power or larger models into general intelligence. Other constraints include unreliable long-horizon execution, limited real-world feedback, evaluation weaknesses, cost and the difficulty of defining a stable milestone.
The pace could be faster or slower than recent benchmark trends suggest. Forecasts should spell out which scientific advances they assume and how they would be validated. Predictions with incompatible definitions cannot be directly averaged into a single objectively established AGI date.
Cortist’s F-001 v0.1 assigns a provisional 38% cumulative chance by 31 December 2030 to its intended AGI threshold. This is an editorial judgment, not a scientific frequency or a scored estimate. The year selector displays historical subjective model values rather than observations. The underlying event criteria are still under review.
An external scientist’s predicted arrival year or a forecasting-platform consensus is not the same as Cortist’s own F-001 record. We keep publication dates, event definitions and source attribution separate. A later revision must be visibly dated; it should not overwrite the earlier curve or retrospectively claim a verified outcome. Read the full F-001 record and forecast protocol.
The strongest positive signal would be reproducible, independently assessed general performance across unfamiliar domains, sustained task completion and credible comparisons with humans on predeclared tasks. Important counterevidence includes failures to generalize, evaluation contamination, dependence on extensive human scaffolding and results that outside evaluators cannot replicate.
A useful report should state not just the success rate but what failed, which tools were allowed and whether the system could detect and correct its own errors. The Cortist methodology is designed to keep claims, provisional forecasts and later real-world verification distinct.
Broad AI capability may have substantial implications for employment, science and governance, but those outcomes also depend on implementation, regulation, costs and human choices. A technically capable model is not proof that an occupation will disappear, an AI-discovered treatment will be approved, or a future system will escape oversight. Those questions require their own evidence.
Cortist follows the consequences separately through AI & Jobs, AI in Medicine, AI Agents, Superintelligence and AI Existential Risk. Separating these topics avoids turning a contested technical milestone into an unsupported prediction about society.
The Levels of AGI framework is valuable because it distinguishes breadth and depth of capability. The framework does not appoint an official authority for deciding whether AGI has arrived. A claim based on benchmark success still requires evidence of transfer to unfamiliar work and an evaluation that outside researchers can reproduce.
METR task horizons describe performance on a defined set of technical tasks at a specified success rate. They cannot by themselves establish general intelligence across all domains. Cortist’s 38% cumulative F-001 view for the end of 2030 remains a provisional editorial forecast, not a scientific measurement.
These publications address different aspects of intelligence and its measurement. None is a certification that AGI has already occurred.
For dated researcher statements and the original forecast record, continue to the expert chronology and forecast sections below. Original model versions and Ledger evidence remain separately preserved.
We separate an expert's publicly stated timeline from numerical probabilities. We never invent a percentage for someone who only gave a date range.
The Google DeepMind leader expects AGI within two to three years, compared with a three-to-five-year estimate earlier in 2026. He describes AGI as covering the full range of human cognitive capabilities.
A long-running forecasting question records changing community expectations for the first general AI system to be devised, tested and publicly announced. The estimate is dynamic, not a fixed statement by one expert.
10 dated source records, 4 years. Direct original references; no new news articles.
A chronological research and public-statement archive, not a leaderboard. Every entry identifies the speaker, publication date, location, core claim and what the statement does not prove. Summaries are editorial paraphrases, not invented direct quotes.
DeepMind proposed a taxonomy of ten cognitive abilities and a human-baseline evaluation protocol for measuring progress toward AGI.
Context: Measurement proposal, not an arrival prediction or proof of AGI.
Read original source ↗The authors argued that high benchmark performance should not be confused with flexible, real-world general intelligence.
Context: Explicit skeptical counterpoint; this is their interpretation, not an independently resolved AGI outcome.
Read original source ↗Marcus predicted that AGI would not arrive in 2026 or 2027 and argued that current LLM systems remained insufficiently reliable.
Context: A dated negative forecast; no numerical probability or shared AGI test specified.
Read original source ↗LeCun discussed limitations of current LLMs, the need for world understanding, abstract reasoning and planning, and a longer research path toward human-level AI.
Context: Recorded interview and methodological position; not evidence that AGI has arrived.
Read original source ↗The authors described human-level capability on most cognitive tasks as potentially arriving in the coming years and called for proactive testing and safety planning.
Context: Broad qualitative time horizon, not a date-certain prediction or a numeric probability.
Read original source ↗METR reported an approximately seven-month historical doubling time for the length of tasks generalist agents could complete at a specified reliability level, while noting extrapolation limits.
Context: Measures a particular autonomous-task capability, not AGI itself.
Read original source ↗Altman expressed confidence in a route to traditionally understood AGI and expected the first AI agents to begin materially affecting workplaces in 2025.
Context: A first-person capability outlook, not a publicly independently validated AGI milestone.
Read original source ↗Amodei explored the potential upsides of powerful AI and repeatedly cautioned that predictions about its consequences involve considerable uncertainty.
Context: His defined powerful AI is not automatically equivalent to Cortist's proposed AGI threshold.
Read original source ↗Researchers proposed levels that separate an AI system's generality from how proficiently it performs, rather than treating AGI as a single simple switch.
Context: Academic measurement framework; no claim that a specific model had reached full AGI.
Read original source ↗OpenAI described AGI as a possible transformative but risky capability and called for safety evaluations, public scrutiny and governance while acknowledging uncertain timelines.
Context: Institutional policy statement, not a probability forecast or achieved milestone.
Read original source ↗Every relevant AGI development belongs in context. News may update the evidence without silently changing Cortist's original forecast.
The deadline is shifting closer. The definition—and the verification problem—still matter.
Read the analysis ↗RESEARCH · 17 MAR 2026A cognitive evaluation framework points to the hard work of testing broad ability, not just isolated scores.
Original research announcement ↗CORTIST FORECAST · OCT 2026The early probability is visible, and the missing resolution requirements are documented rather than hidden.
Inspect the forecast ↗Calling a system AGI is easy. Demonstrating general intelligence without moving the goalposts is harder.
Performance across diverse cognitive domains—not just one benchmark or profession.
Generalization and adaptation to genuinely unfamiliar challenges.
Replicable evaluations beyond a company's own announcement.
A fixed resolution standard, published before the estimate is officially scored.
Editorial transparency: these are principles, not yet a complete pass/fail threshold. Our 38% is provisional until a measurable resolution protocol is locked and time-stamped. See the Cortist methodology and forecast record.
Artificial general intelligence. The term usually refers to AI with broad, transferable intellectual capabilities across tasks and domains. There is no single universally agreed technical test.
That remains unknown. Cortist's current working model v0.1 estimate is 38% by 31 December 2030, but our full resolution protocol is still being finalized. This is not a measured frequency or a guaranteed prediction.
No numerical probability is reported in his 21 October 2026 interview. He gave a two-to-three-year timeline; Cortist does not translate that statement into an invented percentage.
Cortist runs a daily editorial search for newly published or revised AGI predictions. Any new entry is subject to source checks. Page updates and the permanent forecast record are separate from the daily research check.
All forecasts are timestamped, carefully sourced and revisited when the evidence changes.
Explore the evidence on existential risk, loss of control and expert disagreement. The X-001 research question is explicitly unscored.
Read AI Existential Risk ↗News is linked to this research topic, but no article alone proves the associated forecast.
Google DeepMind's leader expects AGI within two to three years. That is an expert timeline, not a measured probability or a verified result.
Read full news report ↗
Google is pitching its new model for complex coding and organizational tasks. The useful measure is not a demo—it is completed work without rescue.
Read full news report ↗View the AGI news filter or explore the complete seven-question forecast register. The topic classification is editorial and the sources remain dated.
