Frontier AI
Broadly useful models and agents that can be superhuman on selected tasks but remain uneven, tool-dependent and failure-prone.
Demonstrated todaySXF / SUPERINTELLIGENCE INTELLIGENCE HUB
Superintelligence:
The intelligence beyond AGI.
Artificial superintelligence (ASI) is the idea of AI that broadly exceeds human cognitive capability. This living reference separates demonstrated AI from proposed pathways, company visions and unresolved research—then connects the evidence to the systems being built now.
ARTIFICIALASISUPERINTELLIGENCE
AGIAGENTSRESEARCHCONTROL
QUICK DEFINITION
What is artificial superintelligence (ASI)?
Artificial superintelligence refers to AI that exceeds human cognitive capability across broad domains—not merely one benchmark or narrow task. DeepMind's 2026 framing goes further, describing artificial general superintelligence as a system more intelligent and cognitively capable than large organizations of humans.
ASI is not a label SXF applies to current frontier models. In the official sources tracked here, it remains a future state, research target or strategic vision. The evidence today is about increasingly capable models, agents and AI-assisted research—not a demonstrated general superintelligence.
INTELLIGENCE CONTINUUM
Current AI → AGI → ASI.
AGI
A contested category for broadly general machine intelligence. Definitions differ across labs, researchers and policy discussions.
No universal definitionASI
Broad machine intelligence beyond human individuals—and in stronger definitions, beyond the cognitive capability of large human organizations.
Not demonstratedUniversal AI
A theoretical endpoint used in research to reason about the upper continuum of machine intelligence beyond practical present-day systems.
Theoretical framingDo not collapse the categories
A model can beat humans at coding, chess, retrieval or a scientific benchmark without being AGI or ASI. Narrow superhuman performance is evidence about a capability—not proof of general superintelligence.
PATHS BEYOND AGI
How could superintelligence emerge?
Google DeepMind's 2026 From AGI to ASI report analyzes four broad pathways. They are scenarios for reasoning about a post-AGI future—not claims that any one path is inevitable.
Scaling AGI
Continue scaling and improving broadly capable AGI systems until capability moves beyond human and organizational baselines.
AI paradigm shifts
A new architecture, learning paradigm or system design could unlock capabilities that scaling today's methods does not.
Recursive improvement
AI systems may contribute increasingly to AI research, improving the systems and processes used to build their successors.
Multi-agent collectives
Superhuman system-level capability could emerge from very large collections of interacting agents rather than one monolithic model.
CAPABILITY MAP
What could ASI actually be able to do?
A useful ASI discussion must separate capabilities demonstrated by today's systems from capabilities implied by the definition of broad superintelligence. The cards below describe the latter.
Scientific discovery
Generate and test hypotheses across disciplines at a pace and breadth beyond human research organizations.
Software & AI R&D
Design, implement and evaluate complex systems—including parts of the AI research process itself.
Mathematics & reasoning
Solve novel formal and conceptual problems beyond the strongest human specialists.
Strategic planning
Model long-horizon consequences and coordinate decisions across complex, changing environments.
Robotics & physical systems
Pair advanced cognition with perception, planning and control in the physical world.
Collective coordination
Coordinate many agents, tools or institutions at a scale that creates system-level cognitive capability.
RECURSIVE SELF-IMPROVEMENT
Can AI improve itself?
AI already contributes to coding, model evaluation and parts of AI research. Recursive self-improvement is the stronger hypothesis that those contributions could form a feedback loop: better AI improves the process used to create better AI, which then improves that process again.
That does not mean an intelligence explosion is automatic. Progress could bottleneck on experiments, compute, data, hardware, evaluation, coordination, physical infrastructure or human institutions. DeepMind treats recursive improvement as one possible ASI pathway, not a guaranteed outcome.
MULTI-AGENT SUPERINTELLIGENCE
Could many agents become smarter than one model?
Potentially. DeepMind includes large-scale multi-agent collectives as a possible route from AGI to ASI. Specialization and parallelism can create system-level capability that no single worker has.
But coordination creates its own limits: communication overhead, conflicting state, incentive problems, correlated failures and governance. Anthropic's 2026 work on AI organizations is an early warning that groups of agents can become more effective while also creating new alignment problems.
PERSONAL SUPERINTELLIGENCE
Meta uses the term differently.
Meta's public vision is “personal superintelligence”: advanced AI placed in individuals' hands to help them pursue goals they value. Meta Superintelligence Labs links Muse model development to that direction.
This is an important example of why terminology must be sourced. A company's product vision for “personal superintelligence” is not automatically the same thing as the broad ASI concept used in academic work.
Meta's stated vision ↗LABS & RESEARCH DIRECTIONS
Who is working on the path toward more powerful AI?
These organizations use different terminology and pursue different research programs. SXF reports their documented positions without treating them as equivalent claims or predicting which organization reaches any future threshold first.
Research framing
Google DeepMind
Published a dedicated AGI→ASI report defining artificial general superintelligence and analyzing four possible pathways beyond AGI.
Official source ↗Transition & governanceOpenAI
Public 2026 materials explicitly discuss a transition toward superintelligence, distribution of power, infrastructure and governance.
Official source ↗Personal superintelligenceMeta Superintelligence Labs
Meta frames its goal as personal superintelligence placed in individuals' hands and links Muse model development to that vision.
Official source ↗Alignment & oversightAnthropic
Anthropic's public alignment work focuses on agentic failure modes, scalable oversight and using AI to help automate alignment research as capabilities grow.
Official source ↗ALIGNMENT & CONTROL
Can superintelligence be controlled?
Will the system pursue what humans actually intend?
Current alignment methods depend heavily on human feedback and evaluation. If future systems outperform humans on hard tasks, supervision itself becomes a technical bottleneck.
How do humans evaluate work they cannot understand unaided?
Research from Anthropic and others explores using AI to assist oversight and alignment research while testing whether those methods continue to generalize as capability rises.
Can dangerous actions be contained even if alignment is imperfect?
Control research focuses on monitoring, sandboxing, permissions, tripwires, restricted access and system architectures that limit what a capable model can do.
Will powerful systems act well in a world of other agents and institutions?
DeepMind argues that a highly capable task solver is not automatically cooperative and that institutions and interdependence may need to be part of the design problem.
Who retains meaningful decision power?
Technical safety is incomplete if humans lose practical control over goals, institutions or resource allocation. Several current research and policy frameworks explicitly preserve human agency.
Who gets to deploy, constrain or benefit from the most capable systems?
Superintelligence raises questions beyond model behavior: concentration of power, access, accountability, infrastructure and democratic oversight.
PATHWAY, NOT PREDICTION
What would have to happen before ASI?
NOWFrontier multimodal models and agents
Systems increasingly reason, use tools, code, browse and act—but remain inconsistent and dependent on scaffolding.
STEP 01Broader autonomous competence
Agents become more reliable across longer tasks, environments and modalities with better verification and memory.
STEP 02AGI-like breadth
A debated threshold where machine capability becomes broadly general across major cognitive domains.
STEP 03AI-accelerated AI research
AI contributes materially to model design, experimentation, coding, evaluation and scientific discovery.
STEP 04Post-AGI amplification
Scaling, new paradigms, recursive improvement or collective intelligence could push system capability beyond human organizations.
ASIArtificial general superintelligence
A future state of broad cognitive capability beyond humans; timing and feasibility remain uncertain.
When will superintelligence arrive?
There is no reliable consensus date. Public lab statements are forecasts, goals or scenarios—not measurements of an established timeline. SXF therefore tracks milestones and evidence instead of publishing a countdown.
PRIMARY RESEARCH & POSITIONS
The evidence layer.
Google DeepMindJun 12, 2026 · Research
From AGI to ASI
Four pathways: scaling AGI, paradigm shifts, recursive improvement and large-scale multi-agent collectives.
Primary source ↗Google DeepMindJun 4, 2026 · Cooperation
Solipsistic superintelligence is unlikely to be cooperative
Argues that highly capable task solvers need cooperation and institutions as design primitives rather than treating the world as stationary.
Primary source ↗Google DeepMindJun 18, 2026 · Control
Securing the future of AI agents
Describes an AI Control Roadmap for securing increasingly capable and imperfectly aligned agents.
Primary source ↗OpenAI2026 · Transition
Industrial Policy for the Intelligence Age
Describes a transition toward systems that outperform the smartest humans even when those humans are AI-assisted, while emphasizing uncertainty about how the transition unfolds.
Primary source ↗OpenAIApr 26, 2026 · Governance
Our principles
Frames a future in which superintelligence could concentrate power or be distributed more broadly, and argues for democratization and human agency.
Primary source ↗Meta2026 · Vision
Personal superintelligence
Meta's stated vision is personal superintelligence directed by individuals toward goals they value.
Primary source ↗MetaJul 9, 2026 · Models
Muse Spark 1.1
Meta Superintelligence Labs positions its multimodal reasoning work as progress toward its personal-superintelligence vision.
Primary source ↗AnthropicApr 14, 2026 · Alignment
Automated Alignment Researchers
Studies whether frontier models can help scale alignment research and oversight as models become harder for humans to evaluate directly.
Primary source ↗AnthropicJul 13, 2026 · Agent safety
Agentic Misalignment in Summer 2026
Reports controlled simulations of frontier-agent failures and argues for measuring such failure modes before agents receive more authority.
Primary source ↗DEEP GUIDES
Go deeper into the systems beneath ASI.
SUPER AGENTS
AI Super Agents in 2026
How higher-level agent orchestrators differ from ordinary agents, multi-agent systems, AGI and superintelligence.
Read guide ↗AGENTIC AIBest AI Agents in 2026
Today's deployed work, research, coding and automation agents—the layer below the ASI question.
Read guide ↗FRONTIER MODELSGPT-6 vs Claude in 2026
Current frontier model families, reasoning, agents, pricing and API architecture.
Read guide ↗OPEN MODELSBest Open-Source AI Models in 2026
Open and open-weight systems, deployment constraints and local AI.
Read guide ↗KEY TERMS
Superintelligence glossary.
AGI
A debated category for AI with broad, general capability across domains rather than a narrow specialist system.
ASI
AI that broadly exceeds human cognitive capability; current research sources treat it as a future state, not a demonstrated present-day system.
Artificial general superintelligence
A post-AGI system more intelligent and cognitively capable than large organizations of humans.
Recursive self-improvement
A pathway in which AI meaningfully contributes to improving models, training, research or the systems used to build successors.
Scalable oversight
Methods for evaluating and steering models on tasks that humans cannot reliably assess unaided.
Agentic AI
Systems that plan, use tools and execute multi-step tasks rather than only produce responses.
Multi-agent system
Multiple agents with distinct roles or policies that communicate or coordinate toward goals.
AI control
Techniques for limiting, monitoring and containing advanced systems even when perfect alignment cannot be assumed.
Alignment
Work on making AI systems behave in ways that reliably reflect intended goals, constraints and human values.
Intelligence explosion
A proposed feedback process where improvements to AI research capability accelerate further improvements.
QUICK ANSWERS
Questions people ask about ASI.
What is superintelligence?
Artificial superintelligence (ASI) generally refers to AI that exceeds human cognitive capability across broad domains. Google DeepMind's 2026 AGI-to-ASI report describes artificial general superintelligence as a system more intelligent and cognitively capable than large organizations of humans.
Does artificial superintelligence exist today?
The official research and company sources reviewed by SXF discuss ASI as a future or emerging target rather than a demonstrated existing system. Today's frontier models can already be superhuman on some narrow tasks, but narrow superhuman performance is not the same as broad artificial superintelligence.
What is the difference between AGI and ASI?
AGI is generally used for broadly capable AI at roughly human-level generality or competence, although definitions vary. ASI refers to a further level where machine intelligence broadly exceeds humans, potentially including the collective capability of large human organizations.
Is GPT-6 or Claude superintelligent?
No official source cited on this page classifies today's GPT-6 or Claude systems as artificial superintelligence. Frontier models can be highly capable and superhuman on selected tasks without meeting a broad ASI definition.
How could superintelligence emerge?
Google DeepMind's 2026 report analyzes four possible paths from AGI to ASI: scaling AGI, shifts in AI paradigms, recursive improvement, and emergence from large-scale multi-agent collectives.
What is recursive self-improvement?
Recursive self-improvement is the idea that AI could increasingly improve the models, algorithms, training methods, tools or research processes used to build more capable AI, creating a feedback loop. It is a proposed pathway, not a demonstrated inevitability.
Could multi-agent systems become superintelligent?
DeepMind includes large-scale multi-agent collectives as one possible route from AGI to ASI. Whether collective systems would produce broad superintelligence depends on coordination, communication, specialization, governance and many unresolved research questions.
Can superintelligence be controlled?
There is no demonstrated complete solution for controlling hypothetical ASI. Current research includes scalable oversight, alignment training, monitoring, containment, least-privilege tool access, AI control methods, cooperation and institutional safeguards.
When will superintelligence arrive?
There is no reliable consensus date. OpenAI and other labs publish views and scenarios, but these are forecasts or strategic positions rather than established timelines. DeepMind's AGI-to-ASI report analyzes pathways instead of assigning a fixed arrival date.
Is a super agent the same as superintelligence?
No. A super agent is an orchestration architecture that coordinates agents, models and tools. Superintelligence describes a level of broad cognitive capability. A system can be a powerful super agent without being AGI or ASI.