The throughline of seventy years isn’t simply bigger models — it’s intelligence moving from answering questions to sustained action, organized collaboration, and finally improving “the methods that make intelligence itself.”
In 2026, “Super Intelligence” shows up in policy, product, and technical discourse at once. Conflating the three mistakes vision for fact.
A September 2026 executive order pushes “Super Intelligence / SI” as the replacement term for AI — while the legal definition still tracks the existing AI scope.
Meta pushes personal superintelligence; SSI makes safe superintelligence its sole mission. Both mix roadmap signaling, brand positioning, and long-term vision.
The classic definition stresses surpassing the best humans across nearly all important cognitive domains; DeepMind goes further, proposing systems that surpass large human organizations.
Superhuman ability in narrow domains ≠ a general agent; a general agent ≠ ASI.
This report uses “SI” for the broad 2026 superintelligence narrative, and reserves “ASI” for Artificial Superintelligence in the technical sense.
Each leap doesn’t simply replace the last stage — it stacks a new capability layer onto the system.
Turing’s Imitation Game reframed “can machines think?” as an evaluable behavioral question.S1
Dartmouth proposed that every aspect of learning and intelligence could in principle be precisely described and simulated by machines.S2
Good observed that designing better machines is itself an intellectual activity — a superintelligent machine might improve the methods that make intelligence.S3
GPUs, data, deep networks, and predictable scaling laws turned chips, power, data centers, and capital into the means of producing capability.S7
Reasoning, tool use, computer operation, and agents stretched the unit of a task from a single answer to hours, days, and beyond.S8
A single model’s “IQ” can’t explain organization-level capability. The real leap comes from coupling models, memory, tools, experiments, resources, and governance.
If a system can reliably coordinate thousands of expert-level agents, it could surpass existing companies, universities, and research institutions at the organizational level — even without one infinitely smart central model.Google DeepMind: From AGI to ASI
Filter by region. Corporate disclosures reflect public roadmaps, not independent third-party verification.
Frontier models, reasoning, agents, computer use, memory, and infrastructure in parallel; automated AI research named one of three top goals for 2026.S9 · S10 · S21
Built on the Alpha series, reinforcement learning, and automated discovery; systematically discusses four post-AGI routes: scaling, paradigm shifts, recursive improvement, and multi-agent collectives.S11 · S22 · S23
The Muse Spark line stresses multimodal reasoning, tools, and multi-agent orchestration — superintelligence as a personal capability amplifier.S12
Its Responsible Scaling Policy maps cyber, bio, AI R&D, and autonomy capability thresholds to escalating safety requirements.S13
No transitional consumer products at the center; an attempt to advance capability and safety in lockstep.S14
Grok Bot as a persistent team member — with identity, memory, tools, a runtime environment, and the ability to collaborate.S15
Seed2.1 is officially described as a participant in the model R&D pipeline, spanning evaluation, data, training, research, and infrastructure.S16
Public narratives rarely say “ASI” outright, but roadmaps converge on long context, agentic coding, cross-tool execution, agentic cloud, and devices.S17 · S18 · S19
The likelier reality isn’t picking one of four — it’s all four stacking and forming feedback loops.
Growing models, data, and training/inference compute keep crossing specialist capability boundaries.
New architectures, world models, memory, neuro-symbolic fusion, or unknown learning mechanisms raise the ceiling of the capability curve.
AI improves models, training, inference harnesses, compilers, chips, and experimental pipelines.
Expert agents, through division of labor, parallelism, and mutual review, form organization-level intelligence beyond any single model.
When AI doesn’t just assist research but starts continuously compressing the next generation’s R&D cycle, the speed of capability growth itself becomes a historical variable.
A one-off efficiency gain is not an intelligence explosion. What counts is continuous, cross-generational, reproducible positive feedback spanning software–models–data–hardware.
The share of research independently or predominantly completed by AI keeps rising at frontier labs.
From paper idea to reproducible experiment — no longer just code snippets and literature reviews.
New algorithms, training methods, or system optimizations produced by AI are steadily used in production systems.
Model R&D cycles shorten markedly, beyond what more headcount alone can explain.
Machine teams can autonomously discover, diagnose, and fix experimental issues, running stably for long stretches.
Simultaneously optimizing software, models, data, inference harnesses, and hardware design.
The Turing test asks “does it seem human?”; the more important question: can a system produce real results exceeding human organizations, over the long term?
Spanning the vast majority of important cognitive domains — not narrow superhumanity.
World-class expert level across multiple domains.
Working stably for weeks or months under low supervision.
Safely invoking software, funds, equipment, and robots.
Coordinating large-scale agents while preserving organizational reliability.
Continuously producing new knowledge beyond training data and literature.
Markedly improving AI R&D speed and success rates.
Improvements that compound across generations and self-reinforce — not one-off efficiency gains.
Opportunity and risk aren’t two separate lists — they’re mirror images of the same capabilities under different goals, permissions, and institutions.
Hypothesis–modeling–experiment–analysis forms a machine loop; bottlenecks shift to compute, experiment automation, and verification.
Strategy, development, law, design, research, and operations get a new supply curve.
Scarce resources shift from executors to good questions, good goals, trusted data, and judgment.
Cheap replication of high-level intelligence could narrow service gaps caused by geography, income, and talent scarcity.
The stronger the capability, the larger the drift from vague goals; overseers may be unable to directly judge a system’s outputs.
Drug discovery, vulnerability discovery, and industrial control capabilities can all produce diametrically opposite outcomes.
The near-term risk may first come from the human organizations holding compute, models, data, and agent infrastructure.
Information generation outruns verification capacity; the distribution of productivity dividends, individual participation rights, and chains of fact all need rebuilding.
True SI won’t just be a smarter chatbot. More likely it’s a distributed, continuously running system able to invoke digital and physical resources.
What humanity should fight for isn’t halting all intelligent progress — it’s keeping capability growth, control mechanisms, social institutions, and human participation evolving in sync.
The first question decides whether we arrive safely; the second decides whether the superintelligence era still belongs to humanity.
This page is a structured analysis of research report V1.1; it doesn’t package corporate visions, policy labels, or future judgments as verified technical facts.