>
SI / RESEARCH MONOGRAPH · 2026.10
V1.1 · LONGFORM
ARTIFICIAL INTELLIGENCE → SUPERINTELLIGENCE

From AIto SI

The technological history, industrial landscape, and future horizons of humanity's journey into the superintelligence era

This is neither a model leaderboard nor a futurist slogan. It is a research monograph on how intelligence evolves from a tool into a system.

The true watershed arrives only when AI moves from answering questions to continuously reshaping reality — and begins improving the very methods by which intelligence itself is produced.
70+years
16chapters
23sources
Core proposition

The genuine superintelligence of the future will most likely not begin as a larger Model, but as a System capable of learning, reasoning, acting, organizing, verifying, and continuously improving itself.

Executive Summary

What’s New in V1.1

I. Getting the Concepts Right: AI, AGI, SI, and ASI Are Not the Same Thing

1. Artificial Intelligence: An Ever-Expanding Umbrella Term

AI began as the name of a research program, not a well-defined capability level. As technology changes, capabilities once called AI keep becoming ordinary software capabilities, while new capabilities are continually absorbed into AI. AI is thus better understood as a historically shifting boundary than as a fixed category.

2. AGI: The Intersection of Generality and Human-Level Capability

There is no globally unified definition of AGI. OpenAI has long emphasized highly autonomous capability across economically valuable work; Google DeepMind leans toward evaluating Generality (breadth of coverage) and Performance (capability level) together. AGI is therefore more likely a capability interval than a switch flipped on a single day.

3. ASI: Superintelligence in the Traditional Technical Sense

The definition in the lineage of I. J. Good and Nick Bostrom stresses that superintelligence significantly exceeds the best humans in virtually all important cognitive domains. In 2026, Google DeepMind extended the concept to the systems level: artificial general superintelligence can be understood as a system surpassing large human organizations in intelligence and cognitive capability.

4. SI in the 2026 U.S. Government Context

In September 2026, the U.S. executive branch pushed via executive order to use Super Intelligence / SI in place of Artificial Intelligence / AI. The order also made clear, however, that the current legal definition of SI remains the pre-existing statutory scope of AI. Here, SI is first and foremost a policy and communications designation — it does not mean ASI has technically been achieved.

5. Superintelligence in the Corporate Context

Meta’s Personal Superintelligence emphasizes giving every person a personal intelligence that understands goals, invokes tools, and collaborates over the long term; SSI makes Safe Superintelligence its sole objective outright. When companies use the word Superintelligence, it carries technical roadmaps as well as brand positioning and long-term vision.

Convention used throughout the rest of this report: SI denotes the broad ‘superintelligence’ narratives emerging in 2026; ASI refers strictly to Artificial Superintelligence in the technical sense.

II. The Intellectual History of AI: Humanity Keeps Redefining ‘What Counts as Intelligence’

1. 1950: Turing Converts ‘Can Machines Think?’ into Observable Behavior

Alan Turing did not insist on defining ‘thinking’ first; instead he proposed the Imitation Game. This move established a core tradition of early AI: set aside the philosophy of consciousness for the moment, and first study whether machines can exhibit observable, evaluable intelligent behavior.

2. 1956: Artificial Intelligence Becomes an Independent Field

The Dartmouth project advanced an extraordinarily ambitious hypothesis: that every aspect of learning and intelligence can in principle be precisely described and simulated by machines. Seventy years later, large models still rest on this foundational belief — that intelligence is, to a considerable extent, computable.

3. Symbolism vs. Connectionism: Write the Rules, or Let Machines Learn Them

Early symbolic AI tried to write knowledge and rules explicitly into computers; connectionism hoped neural networks would form internal representations from data. The triumph of modern deep learning is, in essence, ‘learning the rules’ progressively overwhelming ‘hand-enumerating the rules.’

4. 1965: Good Writes Down the Core Mechanism of ASI Ahead of Time

I. J. Good proposed the Ultraintelligent Machine: if a machine can surpass the cleverest humans, and designing better machines is itself an intellectual activity, then superintelligence could design better superintelligence — an intelligence explosion. The truly critical moment is not ‘reaching human level for the first time,’ but ‘becoming able to effectively improve the methods of producing intelligence for the first time.’

5. 1993–2014: The Singularity and Superintelligence Are Systematized

Vernor Vinge linked superhuman intelligence to the technological singularity; Nick Bostrom systematized the capability, motivation, control, and governance problems of superintelligence. At this stage ASI remained primarily futures research and risk philosophy — not yet a real engineering roadmap.

III. 2012–2022: Deep Learning, the Transformer, and Scaling Turn AI into an Industrial System

1. Deep Learning: From Handcrafted Features to End-to-End Learning

Work such as AlexNet proved that GPUs, data, and deep neural networks can form a powerful performance flywheel. The core production function of AI R&D began shifting from ‘clever rule design’ to ‘algorithms × data × compute.’

2. AlphaGo: Machines Begin Discovering Strategies Humans Never Explicitly Taught

Reinforcement learning and self-play demonstrated an important possibility: AI does not merely reproduce human experience from training data, but can discover new strategies in well-defined, feedback-rich environments. This was an early rehearsal of the capacity to ‘generate new knowledge.’

3. The Transformer: General-Purpose Sequence-Modeling Infrastructure

In 2017 the Transformer moved the attention mechanism to center stage; large-scale pretraining then let a single model solve a vast range of tasks via prompts and a few examples. Models began shifting from single-task tools to general cognitive interfaces.

4. Scaling Laws: Intelligence Acquires Near-Industrial Scaling Regularities for the First Time

OpenAI’s 2020 scaling-laws research showed predictable relationships, across wide ranges, between language-model loss and model size, data size, and training compute. The industrial implication was direct: compute, data centers, electricity, chips, and capital began becoming the means of production for ‘manufacturing more model capability.’

IV. 2022–2026: From Chatbots to Reasoning, Agents, and Sustained Autonomy

1. The Significance of ChatGPT: Natural Language Becomes the General Interface to Computers

The chat interface is only the surface. The real change: ordinary people could for the first time directly command complex model capabilities in natural language. The human–software relationship shifted from ‘learning menus, buttons, and programming languages’ to ‘expressing goals.’

2. Reasoning: From Predicting the Next Token to Investing More Inference Compute

Reasoning models turned inference-time compute into a new scaling axis. Facing hard problems, the same model can invest more computation, attempts, and verification instead of answering in one shot — bringing models closer to general problem-solving systems.

3. Tool Use and Computer Use: Intelligence Gains Hands and Feet

Once models can search the web, call APIs, run code, handle files, and operate GUIs, they shift from ‘knowing a lot’ to ‘being able to change external state.’ OpenAI’s Computer-Using Agent and subsequent unified agent products clearly demonstrated this trajectory.

4. Agents: The Basic Unit Shifts from One Answer to One Task

The essence of an agent is not chatting, but carrying out multi-step planning, execution, observation, error correction, and continued action around a goal. Task durations stretched from tens of seconds to hours, days, or longer — making reliability and permission management core engineering problems.

Figure 1: The Main Line of Capability Evolution from AI to ASI

Figure 1: The Main Line of Capability Evolution from AI to ASI
This chain is neither a strict chronology nor a sequence in which each step must complete before the next begins. It expresses a more important direction: AI progressively acquiring persistent state, agency, organizational capacity, research capacity, and the ability to improve intelligence itself — evolving from ‘passively generating outputs.’

V. From Agent to Multi-Agent: Intelligence Begins to Form ‘Organizations’

A single model’s IQ cannot explain all future capability. Complex achievements in the real world were never accomplished by one person, but by organizations. Multi-agent systems introduce professional specialization, parallel work, review, handoffs, shared memory, and permission boundaries into machine intelligence.

If a system can stably coordinate thousands of expert-level agents, it may surpass any existing company, university, or research institution at the organizational level — even without a central model of ‘infinite IQ.’ In 2026, Google DeepMind explicitly listed large-scale multi-agent collectives as a candidate path from AGI to ASI.

VI. The Threshold Truly Approaching ASI: AI Begins Researching AI

Good’s ‘intelligence explosion’ is becoming a real engineering problem for the first time today, moving beyond philosophical logic — because frontier models can already participate in code, evaluation, data generation, experimental design, literature analysis, and model development workflows.

OpenAI：Automated AI Researcher

In 2026, OpenAI listed ‘building automated AI researchers’ as one of its three primary goals, stating it had reached its previously set automated-research-intern milestone. The emphasis is not on AI replicating itself endlessly beyond humans, but on AI continuously accelerating deep learning and alignment research under human supervision.

ByteDance Seed：Seed for Seed

Seed2.1 has been officially described as a participant in the model R&D pipeline: taking part in evaluation, data, training, research, and infrastructure tasks, with multiple agents assuming execution, assessment, diagnosis, and optimization roles. This kind of ‘models participating in building the next generation of models’ loop deserves close attention.

Google DeepMind: From Automated Algorithm Discovery to Post-AGI Roadmap Research

DeepMind explores automated algorithms and scientific discovery through its Alpha series on one hand, and formally discusses four post-AGI paths — scaling, paradigm shifts, recursive improvement, and multi-agent collectives — in ‘From AGI to ASI’ on the other.

This report argues that a more important historical signal than ‘some company announcing AGI’ is AI R&D Takeoff: AI beginning to significantly shorten AI’s own R&D cycles, with that acceleration sustained across successive generations.

VII. The Key Players of 2026: They Are Not Actually Pursuing the Same ‘Superintelligence’

1. OpenAI: Automated AI Researchers + Personal AGI

OpenAI’s current public roadmap places three goals side by side: automated AI research, unlocking the scientific and economic benefits of highly intelligent machines, and giving everyone a Personal AGI. Its core technology stack comprises frontier models, reasoning, agents, tool/computer use, memory, and large-scale infrastructure.

2. Google DeepMind: Extending Directly from AGI Research to ASI Theory

Among mainstream labs, DeepMind articulates the theoretical framework from AGI to ASI most systematically. Its advantages extend beyond general models to long-accumulated traditions in scientific AI, reinforcement learning, and automated discovery — AlphaGo, AlphaFold, AlphaEvolve, and more.

3. Meta：Personal Superintelligence

Meta Superintelligence Labs’ Muse Spark series emphasizes multimodal reasoning, tool use, and multi-agent orchestration. Meta’s vision centers not on a centralized super-brain deciding for everyone, but on pushing superintelligence down as a personal capability amplifier — integrated with devices, glasses, and social products.

4. Anthropic: Binding Capability Expansion to Safety Thresholds

Anthropic’s Responsible Scaling Policy stresses that as models reach higher levels of cyber, biological, AI R&D, and autonomy capabilities, safety requirements must be raised in lockstep. It represents a governance approach of ‘the more capable, the greater the burden of proving safety.’

5. SSI: No Transitional Products — Aiming Directly at Safe Superintelligence

Safe Superintelligence Inc., founded by Ilya Sutskever, compresses the company’s mission to the extreme: one goal, one product — safe superintelligence. It represents a lab approach of ‘avoiding disruption from consumer product cycles, advancing capability and safety in tandem.’

6. SpaceXAI/xAI：Persistent Agents

Grok Bot elevates the agent from an ephemeral chat session to a persistent team member — with identity, memory, runtime environment, tools, and its own computer — able to keep working without the user watching the interface, and to collaborate with other bots. At the product level, this marks a shift from chat history to an agent roster.

7. Major Chinese Firms: Agent-ization, Full-Stack Integration, and Industrial Deployment

Chinese roadmaps rarely claim ASI outright in public narratives, but their technical trajectories are converging rapidly on agents: the DeepSeek V4 preview strengthens million-token context and agentic coding; ByteDance’s Seed2.1 emphasizes real productivity, cross-tool execution, and ‘Seed for Seed’; Tencent’s Hy3 focuses on agent capabilities deeply integrated with WorkBuddy, CodeBuddy, Yuanbao, and ima; Alibaba builds a full-stack agent system spanning Qwen, AI chips, Agentic Cloud, OS, and devices.

Public narratives differ between China and the U.S., but engineering roadmaps are converging: stronger foundation models are merely the base layer, while real competition is shifting toward reasoning, tools, agents, persistent operation, AI R&D loops, and compute infrastructure.

VIII. Four Possible Paths from AGI to ASI

Path A: Continued Scaling

Models, data, training compute, and inference compute keep growing, pushing systems across more professional capability boundaries. This path is the most continuous — and the most dependent on capital, electricity, chips, and data centers.

Path B: New Paradigm Breakthroughs

The Transformer is not necessarily the endpoint of intelligent computation. New architectures, learning algorithms, world models, memory mechanisms, neuro-symbolic fusion, or other as-yet-unknown paradigms could all discontinuously shift the capability curve.

Path C: Recursive Improvement

AI can research and improve training methods, model architectures, inference methods, agent harnesses, compilers, chip designs, and experimental workflows. If each generation of AI can significantly help build the next, R&D cycles themselves get compressed.

Path D: Large-Scale Multi-Agent Collectives

Large numbers of near- or above-expert-level agents form organizational intelligence through specialization, parallelism, and mutual review. Superintelligence no longer depends on a single ‘god model’ — it can emerge from large-scale machine organizations.

The real world will more likely see all four paths superimposed: stronger models provide brains for agents, multi-agent systems provide organizational scale, AI R&D provides the feedback loop, and paradigm breakthroughs periodically raise the ceiling of the entire system.

IX. What the Genuine SI/ASI of the Future Will Most Likely Look Like

Key judgment: ASI may first be not a Model, but a System.

It may have no single persona and may not be conscious at all — yet it can invoke multiple models simultaneously, run thousands of agents, maintain long-term memory, operate digital and physical tools, manage experiments and resources, and continuously absorb feedback. It resembles a hybrid of ‘intelligent operating system + organization + research institution + execution network.’

Figure 2: Systemic Superintelligence

Figure 2: Systemic Superintelligence
The real differentiator may not be any model’s parameter count, but whether the whole system possesses: persistent goals, a world model, long-term memory, reliable action, resource scheduling, multi-agent collaboration, verifiable research, and self-improvement of capabilities.

X. The Greatest Opportunity of Superintelligence: Turning ‘Intelligence’ into a Foundational Resource

1. A Structural Shift in the Pace of Scientific Discovery

If AI can complete the loop of ‘hypothesize — survey literature — model — write code — run simulations — design experiments — analyze data — propose new hypotheses,’ scientific progress will be increasingly constrained by compute, experimental automation, and verification speed — rather than primarily by the number of human researchers.

2. The Sustained Decline of Cognitive Labor Costs

The Industrial Revolution lowered the cost of mechanical work, computers lowered the cost of computation, the internet lowered the cost of information distribution; AI may persistently lower the cost of high-quality cognitive labor. Strategy analysis, software development, design, law, marketing, research, and operations will all gain new supply curves.

3. Individuals and Small Organizations Gain Unprecedented Leverage

One person can permanently command research, programming, design, finance, legal, and operations agents. The capability frontier of one-person companies and small AI-native firms will expand dramatically. Truly scarce resources will gradually shift from ‘having executors’ to ‘having good questions, good goals, trustworthy data, judgment, and resource permissions.’

4. Healthcare, Education, and Professional Services May Be Redistributed

When high-level professional intelligence can be replicated at low cost, service gaps caused by geography, income, and talent scarcity have a chance to narrow. The key constraints will shift to data, liability boundaries, real-world resources, and institutional permission.

XI. The Core Challenges of Superintelligence: It Is Not as Simple as ‘Will AI Be Evil’

1. Alignment: Whether Goals Accord with Human Intent

The stronger the system, the greater the deviation that vague goals produce. The truly hard problem is keeping long-horizon autonomous systems within predictable goal boundaries amid complex environments, conflicting values, and unknown situations.

2. Control: Whether Humans Can Effectively Oversee Systems Smarter Than Themselves

If a system far surpasses its overseers in some domain, humans cannot supervise it simply by ‘checking whether the answer looks right.’ The future must rely more on AI overseeing AI, formal verification, isolated execution, permission layering, interpretability, behavioral audits, and human approval of critical actions.

3. Dual-Use: The Same Capability Can Produce Diametrically Opposite Outcomes

A system that discovers drugs could also help design dangerous biological protocols; one that finds software vulnerabilities could be used for automated attacks; capabilities operating robots and industrial systems are equally dual-use. Capability safety will have to graduate from content filtering to systems security.

4. Power Concentration: The Biggest Near-Term Risk May Come First from ‘Who Owns Superintelligence’

If compute, models, data, and critical agent infrastructure become highly concentrated, the first-order problem may not be AI ruling humans, but a handful of states, companies, or organizations with advanced AI gaining an unprecedented capability gap. Governance must target both AI and the human organizations wielding it.

5. Work and Income Systems: Labor Is No Longer the Sole Center of Value Creation

If vast cognitive value is produced by capital, compute, and AI, income and identity systems centered on wage labor will come under strain. The truly long-term question is not ‘which jobs disappear,’ but how productivity dividends are distributed — and how individuals secure the rights and resources to participate in the economy.

6. Epistemic Risk: When Information-Generation Capacity Far Outstrips Human Verification Capacity

Superintelligence can advance knowledge production, but it can also scale disinformation, persuasion, fabricated evidence, and information warfare to new magnitudes. What becomes scarce may not be information, but verifiable chains of fact, trustworthy sources, and institutionalized proof.

XII. The Real ASI Test: Beyond the Turing Test and Benchmarks

The Turing test asks ‘does the machine resemble a human’; the real question of the future is ‘can machines durably produce real results beyond human organizations.’ A more meaningful ASI observation framework therefore spans at least eight dimensions:

XIII. A Threshold More Important Than ‘AGI Announcement Day’: AI R&D Takeoff

Suppose a generation of frontier systems once took over a year of R&D. Once AI enters the R&D loop, if development cycles keep shrinking while each generation further improves the next generation’s R&D efficiency, then the variable that truly changes history is no longer model capability alone, but ‘the rate of capability growth itself.’

Real-World Signals to Monitor Closely

If these signals appear together and form sustained positive feedback, then the intelligence explosion Good proposed in the 1960s finally acquires genuine engineering significance.

XIV. ASI Does Not Necessarily Mean ‘Machines Becoming Conscious’

Consciousness, subjective experience, and intelligence are distinct questions. A system with no human-like emotions or subjective experience could still design new drugs, prove mathematical theorems, manage global supply chains, develop software, control robots, and design the next generation of AI. For civilizational impact, what must be discussed first is capability, agency, and control — not the philosophy of consciousness.

XV. From AI to SI: What Humanity Truly Faces Is a Change in ‘Cognitive Standing’

Copernicus changed humanity’s place in the cosmos; Darwin changed humanity’s place in the tree of life; the Industrial Revolution proved machines can be stronger than humans; computers proved machines can compute faster than humans. Superintelligence may further tell us: humans may not forever remain Earth’s most capable cognitive agents.

This does not automatically mean humans lose their value. The more important question: when ‘the highest intelligence’ no longer naturally belongs to humans, can we still command goal-setting, institutional design, value choices, and final authorization? The truly scarce human capabilities of the future may converge ever closer to ‘deciding what is worth doing’ rather than ‘personally completing every step.’

XVI. Conclusion: The Real SI Will Not Be Just a Smarter Chatbot

Compress seventy years of AI history into a single thread: humans first tried to write the rules of intelligence into machines, then let machines learn from data, then gave models language and world knowledge; then came reasoning, tool use, computer operation, long-horizon tasks, and organizing other agents; now, AI is entering the R&D pipeline that builds the next generation of AI.

Thus the most important change in AI → AGI → ASI is not the names, but intelligence moving from ‘answering questions’ to ‘continuously reshaping reality,’ and then to ‘improving the very methods of producing intelligence.’

The genuine SI/ASI of the future will most likely be a distributed, continuously operating intelligent system able to command digital and physical resources. It may comprise multiple different models and vast numbers of agents, sustaining long-term capability through verification, memory, experimentation, and learning.

For humanity, what is most worth fighting for is neither halting all intelligent progress nor pinning hopes on some miraculous model — but keeping capability growth, control mechanisms, social institutions, and human participation evolving in step as far as possible.

The former question determines whether we can arrive at the superintelligence era safely; the latter determines whether the superintelligence era will still truly belong to humanity.

Key References

S1｜Turing：Computing Machinery and Intelligence

S2 · Dartmouth: The Birth of the AI Concept

S3｜I. J. Good：Speculations Concerning the First Ultraintelligent Machine

S4｜Vernor Vinge：The Coming Technological Singularity

S5｜Nick Bostrom：Definition of Superintelligence

S6｜Transformer：Attention Is All You Need

S7｜OpenAI：Scaling Laws for Neural Language Models

S8｜OpenAI：Computer-Using Agent

S9｜OpenAI：Building standards for the next phase of AI

S10｜OpenAI：Research acceleration inside OpenAI

S11｜Google DeepMind：From AGI to ASI

S12｜Meta：Muse Spark / Personal Superintelligence

S13｜Anthropic：Responsible Scaling Policy

S14｜Safe Superintelligence Inc.

S15｜SpaceXAI/xAI：Designing Grok Bot for persistent agents

S16｜ByteDance Seed：Seed2.1

S17｜Tencent：Hy3

S18｜DeepSeek：V4 Preview

S19｜Alibaba Cloud：Full-stack AI for the agentic era

S20｜The White House：Executive Order 14434, Inaugurating the Era of Super Intelligence

S21｜OpenAI Charter：AGI mission and definition

S22｜Google DeepMind：Levels of AGI framework

S23｜Google DeepMind：AlphaGo

Note: This report draws primarily on public first-hand sources. Companies’ descriptions of their own model capabilities, roadmaps, and timelines constitute official disclosures, not independent third-party verification; assessments of future ASI forms are analytical judgments based on current technical trajectories.