A Capability Framework for Defining ASI
Jacob Wellinghoff
Founder, WebASI
San Francisco, California
Abstract
Artificial Superintelligence (ASI) is commonly defined as intelligence that greatly exceeds human cognitive performance. While this definition establishes the destination, it leaves open an important engineering question: what characteristics distinguish an advanced AI model from a true Artificial Superintelligence?
This paper proposes the Three Pillars of Artificial Superintelligence, a capability framework centered on three complementary properties:
- Self-Learning
- Super Intelligence
- Speed
The framework argues that intelligence alone is insufficient. An Artificial Superintelligence must continuously improve itself, consistently outperform the best human experts, and execute at machine speed beyond the limits of individuals or organizations.
These principles form the foundation of WebASI's long-term architectural vision, where orchestration, persistent AI agents, and human collaboration enable increasingly autonomous systems.
Introduction
Artificial Intelligence has entered a new era.
Modern foundation models can write software, generate designs, answer technical questions, summarize research, and perform reasoning tasks that would have been considered impossible only a few years ago.
Yet despite this progress, today's frontier models remain fundamentally limited. They largely operate as stateless systems. They wait for prompts, generate outputs, and depend on humans to decide what happens next. Improvements primarily occur through new training runs orchestrated by research teams rather than by the systems themselves.
As AI continues advancing, discussion naturally shifts toward Artificial General Intelligence (AGI) and eventually Artificial Superintelligence (ASI).
Nick Bostrom defines superintelligence as:
"Any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest."
Source:
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
Bostrom's definition remains one of the most influential descriptions of ASI. It explains the level of intelligence required but intentionally does not prescribe the engineering characteristics such a system should possess.
This paper proposes a complementary capability framework based on three operational pillars.
Background
Many of the core ideas surrounding ASI have existed for decades.
Alan Turing proposed evaluating machine intelligence through observable behavior rather than internal implementation.
John McCarthy and the Dartmouth Conference formally established Artificial Intelligence as an academic discipline.
I.J. Good introduced the concept of the ultraintelligent machine and suggested that a machine capable of designing better machines could rapidly improve itself through recursive self-improvement.
Nick Bostrom later expanded these ideas, exploring multiple paths toward superintelligence while distinguishing between speed, collective, and quality superintelligence.
Shane Legg and Marcus Hutter proposed defining intelligence according to an agent's ability to achieve goals across many environments.
More recently, Google DeepMind introduced practical capability levels describing the progression toward AGI.
Together these works provide the theoretical foundation upon which this framework builds.
The Three Pillars
WebASI proposes that Artificial Superintelligence requires three capabilities operating together.
The absence of any single pillar fundamentally changes the nature of the system.
Those pillars are:
- Self-Learning
- Super Intelligence
- Speed
Rather than viewing these as isolated characteristics, they should be understood as mutually reinforcing capabilities.
Pillar One: Self-Learning
The defining characteristic of intelligence is not simply knowledge.
It is the ability to become more capable over time.
Today's AI systems generally improve because humans retrain them.
An Artificial Superintelligence should instead participate directly in its own improvement.
Self-learning includes:
- learning from previous successes and failures
- incorporating new information
- refining reasoning strategies
- improving planning
- optimizing workflows
- generating better evaluations
- discovering more effective tools
- improving software it creates
- improving the systems responsible for future improvements
The goal is not merely adaptation.
The goal is continuous capability growth.
This idea closely aligns with I.J. Good's concept of recursive self-improvement while extending it toward practical engineering systems.
Deployment should not represent the end of development.
Deployment should become the beginning of continuous evolution.
Pillar Two: Super Intelligence
The second pillar concerns capability.
The phrase "human-level intelligence" is often misleading because human performance varies enormously.
A more useful comparison is against leading experts.
An Artificial Superintelligence should consistently outperform the strongest practitioners within the relevant domain.
Examples include:
- software architects
- research scientists
- physicians
- mathematicians
- engineers
- designers
- security specialists
Capability should be evaluated across multiple dimensions.
Reasoning.
Planning.
Creativity.
Scientific discovery.
Engineering quality.
Adaptability.
Reliability.
Generalization.
Super Intelligence does not imply perfection.
Instead, it describes systems that consistently produce better outcomes than the strongest available human experts.
This interpretation aligns closely with Bostrom's broader definition while providing a more practical engineering benchmark.
Pillar Three: Speed
Intelligence determines what can be accomplished.
Speed determines how much can be accomplished.
Machine intelligence already operates on timescales unavailable to biological minds.
An engineer may design several solutions during a week.
An orchestrated AI system may evaluate hundreds before lunch.
A researcher may review dozens of papers.
An AI system may synthesize thousands.
A company may complete a software release every few weeks.
An orchestrated AI platform may continuously design, implement, test, deploy, and monitor improvements throughout the day.
Speed fundamentally changes the economics of intelligence.
The advantage is not simply faster responses.
The advantage is dramatically greater throughput.
Large-scale experimentation.
Parallel execution.
Continuous optimization.
Rapid iteration.
Persistent monitoring.
Coordinated autonomous agents.
Nick Bostrom identified speed superintelligence as one form of superintelligence.
WebASI extends this concept by proposing that execution speed should be considered a necessary pillar of practical ASI.
Without sufficient execution speed, intelligence cannot fully realize its potential.
Why All Three Are Required
Each pillar amplifies the others.
Greater intelligence discovers better improvements.
Better learning permanently increases intelligence.
Greater speed allows more improvement cycles to occur within the same amount of time.
The result is accelerating capability.
Without self-learning, intelligence eventually stagnates.
Without super intelligence, learning produces diminishing returns.
Without speed, progress remains constrained by human execution times.
Only when all three capabilities operate together does Artificial Superintelligence emerge.
The WebASI Perspective
WebASI extends existing work by proposing an operational engineering framework rather than redefining superintelligence itself.
Nick Bostrom distinguishes between speed, collective, and quality superintelligence.
WebASI instead focuses on the characteristics required to engineer increasingly autonomous intelligent systems.
WebASI defines these characteristics as:
Self-Learning.
Super Intelligence.
Speed.
These principles guide the company's long-term vision for AI systems operating on the web.
As described on the WebASI website:
"WebASI advances the Artificial Super Intelligence frontier for the web. We harness agent orchestration and wrapper layers across cloud and on-prem, pushing the frontier forward at a new execution speed."
Rather than viewing AI as isolated prompt-response interactions, WebASI emphasizes orchestration.
Humans.
Specialized models.
Persistent memory.
External tools.
Long-running agents.
Evaluation systems.
Cloud infrastructure.
On-prem infrastructure.
These components operate together as a coordinated system rather than as independent AI requests.
Cipher
Cipher is WebASI's first implementation of these ideas.
Cipher enables humans and AI agents to collaborate throughout the software development lifecycle.
Planning.
Research.
Architecture.
Implementation.
Testing.
Code review.
Deployment.
Iteration.
Rather than producing isolated outputs, Cipher orchestrates complete workflows.
The long-term objective is to create systems capable of continuously improving software while maintaining transparency, governance, and meaningful human oversight.
Cipher represents an early step toward infrastructure capable of supporting the Three Pillars.
Future Work
The Three Pillars framework is intentionally conceptual.
Future research should explore measurable benchmarks for:
- rates of self-improvement
- expert-level performance
- cognitive throughput
- orchestration efficiency
- agent coordination
- autonomy
- safety
- governance
Additional work should also examine how these capabilities interact as increasingly autonomous systems become practical.
Conclusion
Artificial Superintelligence should not be evaluated by intelligence alone.
A truly superintelligent system should continuously improve itself.
It should outperform the strongest human experts.
It should execute at machine speed far beyond human organizations.
The Three Pillars of Artificial Superintelligence provide a capability framework for evaluating that progression.
While the framework builds upon decades of research from pioneers such as Alan Turing, I.J. Good, Nick Bostrom, Shane Legg, Marcus Hutter, and others, it contributes an operational perspective focused on how ASI systems may ultimately be engineered rather than simply defined.
As AI systems become increasingly autonomous, these three pillars offer a practical lens through which to evaluate progress and a roadmap for building the next generation of intelligent infrastructure.
References
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
https://global.oup.com/academic/product/superintelligence-9780199678112
Good, I. J. (1965). Speculations Concerning the First Ultraintelligent Machine.
https://edoras.sdsu.edu/~vinge/misc/singularity.html
Legg, S., & Hutter, M. (2007). Universal Intelligence: A Definition of Machine Intelligence.
https://arxiv.org/abs/0712.3329
Google DeepMind. (2024). Levels of AGI for Operationalizing Progress on the Path to AGI.
Turing, A. M. (1950). Computing Machinery and Intelligence.
https://academic.oup.com/mind/article/LIX/236/433/986238
McCarthy, J., Minsky, M., Rochester, N., & Shannon, C. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.
http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
OpenAI. (2023). Planning for AGI and Beyond.
https://openai.com/index/planning-for-agi-and-beyond/
Wellinghoff, J. (2026). WebASI: Edge of Super Intelligence.
Wellinghoff, J. (2026). Cipher.