Circle of Life 3.0
Capability is not authority. Intelligence is not permission. Autonomy is not legitimacy. Memory is not truth.
There is a temptation, in 2026, to settle the question of artificial general intelligence with a single number. A benchmark saturates, a chief executive gives an interview, a model ships, and the headline writes itself. This paper takes a different view. Real intelligence is not a score a model earns in isolation. It is something that happens in the world, between people, machines, tools, institutions, memory, and consequences. Circle of Life 3.0 is our attempt to describe that relationship clearly, and to hold current evidence against it honestly.
The thesis, in plain terms
A model on its own is only half of the story. What matters is the whole system it lives inside: the human who sets the purpose and holds the authority, the model that supplies capability and variation, the tools and environments that carry the effects, the institutions that grant or withhold legitimacy, and the memory that preserves what happened. We judge intelligence relationally, by five questions. What can the system do. What is it permitted to do. What does it change. What record does it leave. And can it be undone.
The loop, and the two fields around it
The Circle of Life moves from human purpose and sovereign authority, to synthetic capability and variation, to governed action and effect, to evidence and memory, to a transformed human and institutional state, and then back to renewed purpose. Version 3.0 draws two fields around that loop, because the loop no longer lives only in software. The first is the world and embodiment field: sensors, robots, energy, compute, supply chains, healthcare, mobility, and manufacturing. The second is the institution and jurisdiction field: law, policy, authority, standards, rights, liability, markets, and public legitimacy. Intelligence now leaves marks in both.
Eight axes, and the one that decides
Rather than call a system AGI or not, we measure it across eight axes: breadth, depth, learning, autonomy, embodiment, persistence, relational integration, and governed authority. The first seven describe what a system can do and how it behaves. The eighth is different in kind. Governed authority is about what a system is actually permitted to do, with what oversight, and whether its actions can be reversed. A system can be extraordinarily capable and hold no authority at all. A far narrower system can hold carefully bounded authority. Keeping that distinction visible is the point of the whole framework.
On consciousness we are deliberately restrained. No benchmark discussed here settles whether a machine has experience, moral status, or meaning, and the old line that machines simply do not think does not settle it either. What we can say is operational: current frontier systems reason, generate, perceive, use tools, and increasingly act over long horizons under defined conditions. Human and institutional authority stays explicit regardless of how impressive that becomes.
What OpenAI actually announced with GPT-6 Astra
OpenAI's GPT-6 Astra page, read on 8 September 2026, describes a staged and limited rollout rather than confirmed general availability. It reports strong results on named evaluations, and we keep the evaluation settings attached to each one, because a benchmark is performance on a stated test, not a certificate of general intelligence. Astra is reported at 99.9 percent on ARC-AGI-3, though on a modified responses-API harness and framed by ARC Prize as human parity on the benchmark; 97.6 percent on FrontierMath Tier 4; 96.0 percent on GPQA Diamond; 57.9 percent on Terminal-Bench 4.0; and 72.6 percent on an offline subset of OSWorld 2.0. It is classified Critical for cybersecurity under OpenAI's own Preparedness Framework, and the shipped version refuses advanced offensive tasks. These are vendor-run, vendor-reported figures, and several of the strongest use research configurations that differ from the shipped product. Saturation is not arrival.
NVIDIA, physical AI, and the return of atoms
NVIDIA's message in 2026 is that the frontier is moving out of pure information systems and into the physical world. Jensen Huang has said physical AI has arrived and that every industrial company will become a robotics company, which is a claim about robotics adoption, not a claim about general intelligence. The company and the Japanese government announced national AI infrastructure with 13,750 Vera CPUs, 27,500 Rubin GPUs, and 140 megawatts of power, in support of Japan's stated goal of more than 30 percent of the global AI-robotics market by 2040. That last figure is a government target, not a delivered result. Huang has separately said, in reported interviews, that AGI is achieved for many tasks, using a redefined economic measure; that is his opinion, it is disputed, and we treat it as such.
The definition debate, without a false consensus
There is no settled definition of AGI in 2026, and it would be dishonest to pretend otherwise. Different institutions emphasize different things, and Circle 3.0 keeps each attributed to its author.
| Source | Emphasis | Circle 3.0 reading |
|---|---|---|
| OpenAI Charter | Highly autonomous systems outperforming humans at most economically valuable work; timeline uncertain | Economic breadth and autonomy; it does not settle governance or legitimate authority |
| DeepMind, Levels of AGI (2023) | Performance, generality, and autonomy across staged levels | A useful capability taxonomy; one saturated benchmark does not place a system at a level |
| NVIDIA (Huang) | AGI achieved for many tasks, via a redefined metric | Executive opinion, disputed; attributed to speaker and date |
| Anthropic (OSTP) | Powerful AI in late 2026 or early 2027 | A policy forecast, not a released-system fact |
| ASI discourse | Capability exceeding the best humans across most domains | A scenario class, not a verified state |
Where things actually stand: the 2026 baseline
The Stanford AI Index for 2026 gives a grounded picture, and we keep each figure with its denominator. More than 90 percent of frontier models now come from industry. Reported AI incidents rose to 362, up from 233 the year before. United States private AI investment reached 285.9 billion dollars in 2025, against 12.4 billion in China. Roughly 53 percent of the population reports using generative AI. Experts are far more optimistic than the public about jobs, 73 percent against 23 percent. Some Index figures arrived without clean denominators and we hold those at medium confidence rather than dress them up.
How governments are responding
The most important distinction here is between a law and a plan. The European Union is the only binding law in this set: from 2 August 2026 the AI Office and national authorities began enforcing the AI Act, so systems must disclose when a user is interacting with AI, deepfakes must be labelled, and AI-generated content must carry machine-readable marks. The United States AI Action Plan (July 2025) and the United Kingdom AI Opportunities Action Plan (January 2025, with a pledge to expand research compute at least twentyfold by 2030) are plans, not statutes. China has issued a series of plans and programs across 2026 covering the smart economy, AI cooperation, and AI in the information and communications sector. Japan's METI launched a national program for foundation models in robotics and physical AI. The OECD acts as a tracker across jurisdictions, and the United Nations convened a scientific panel and a global dialogue in July 2026, which are advisory rather than binding.
What this means for work
The honest framing for labor is transformation with real disruption underneath, not a single verdict. The World Economic Forum's 2025 employer survey, covering more than a thousand of the largest employers, projects 170 million new roles and 92 million displaced by 2030, a net gain of 78 million, alongside 22 percent churn in formal jobs. The International Labour Organization finds that about one in four workers are in occupations with some exposure to generative AI, and stresses transformation over wholesale redundancy. The International Monetary Fund estimates about 40 percent of global employment is exposed, higher in advanced economies. Every one of these is a projection or an exposure estimate. Exposure is not the same thing as unemployment.
Where ETHRAEON fits
Everything above points back to the eighth axis. Circle of Life 3.0 is, in the end, an argument for a control plane: authorization at the moment of action, fail-closed gates for consequential and irreversible acts, a tamper-evident evidence trail, persistent memory and continuity, the ability to reverse a decision, and quality assurance across the whole system rather than one component. The decisive unit of analysis is no longer the model by itself. It is the governed system of humans, machines, institutions, and environments, and the questions that matter are who authorized an action, what it touched, what evidence it left, whether it can be undone, and who carries the cost.
Conclusion
Thresholds of capability, product, economics, and politics are being crossed at different moments under different definitions. The honest posture is neither that AGI has definitively arrived nor that it definitely has not. It is to govern intelligence by authority, evidence, continuity, and reversibility, so that whatever does arrive stays human-sovereign. Humanitas Ante Machinam.
Sources, retrieved 8 September 2026, review by 1 March 2027: OpenAI GPT-6 Astra, Charter, and ten-year retrospective; NVIDIA physical-AI and Japan-infrastructure releases, with contemporaneous reporting for spoken statements; DeepMind Levels of AGI (2023); METR time horizons (8 May 2026); Anthropic OSTP recommendations; Stanford AI Index 2026; WEF Future of Jobs 2025; ILO Generative AI and Jobs 2025; IMF (2024); EU AI Act enforcement; US, UK, China, and Japan government sources; OECD; and the UN scientific panel.